- id: INV-001 title: "No explicit ethics module or moral scoring layer." claim_type: invariant subject: ethics.emergence polarity: denies status: active depends_on: [] location: docs/invariants.md#inv-001 source: - docs/processed/legacy_tree/docs/invariants.md - docs/processed/legacy_tree/README.md - id: INV-002 title: "Coherence includes temporal/phase binding, not just static metrics." claim_type: invariant subject: coherence.definition polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-002 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-003 title: "Language emerges as functional self-representation, not a bolt-on." claim_type: invariant subject: language.emergence polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-003 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-004 title: "Post-commit consequence traces are persistent, not resettable." claim_type: invariant subject: consequence.persistence polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-004 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-005 title: "Harm to others contributes via mirror modelling, not symbolic rules." claim_type: invariant subject: ethics.mechanism polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-005 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-006 title: "Post-commit consequence traces cannot be erased, only integrated." claim_type: invariant subject: consequence.non_erasability polarity: denies status: active depends_on: [] location: docs/invariants.md#inv-006 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-007 title: "Language cannot override embodied harm sensing." claim_type: invariant subject: language.constraint polarity: denies status: active depends_on: [] location: docs/invariants.md#inv-007 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-008 title: "Precision is routed and depth-specific, not global." claim_type: invariant subject: precision.routing polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-008 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-009 title: "Attention is precision modulation, not symbolic control." claim_type: invariant subject: attention.mechanism polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-009 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-010 title: "Offline integration exists and is required." claim_type: invariant subject: sleep.necessity polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-010 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-011 title: "Imagination must be possible without belief update." claim_type: invariant subject: default_mode.safety polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-011 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-012 title: "Responsibility arises through commitment, not prediction alone." claim_type: invariant subject: commitment.epistemology polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-012 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-013 title: "Cognition is predictive, iterative, and multi-timescale." claim_type: invariant subject: cognition.structure polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-013 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-014 title: "Representation and regulation are strictly separated." claim_type: invariant subject: architecture.separation polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-014 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-015 title: "Ethics emerges from constraint, not optimisation." claim_type: invariant subject: ethics.emergence_mechanism polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-015 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-016 title: "Stability is prioritized over maximal performance." claim_type: invariant subject: design.priority polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-016 source: - docs/processed/legacy_tree/docs/invariants.md - id: INV-017 title: "Runaway behavior is a control failure, not representational." claim_type: invariant subject: failure.classification polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-017 source: - docs/processed/legacy_tree/docs/invariants.md - id: ARC-001 title: "E1 is the persistent predictive substrate." claim_type: architectural_commitment subject: E1.persistent_predictive_substrate polarity: asserts status: active depends_on: - INV-013 - ARC-004 - ARC-005 location: docs/architecture/e1.md#arc-001 source: - docs/processed/legacy_tree/docs/architecture/e1.md - docs/processed/legacy_tree/architecture/E1.md - id: ARC-002 title: "E2 is the fast forward predictor of affordances." claim_type: architectural_commitment subject: E2.fast_forward_predictor polarity: asserts status: active depends_on: - INV-013 - ARC-004 - ARC-001 - ARC-005 location: docs/architecture/e2.md#arc-002 source: - docs/processed/legacy_tree/docs/architecture/e2.md - docs/processed/legacy_tree/architecture/E2.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-09_e2_hpc_interface.md - id: ARC-003 title: "E3 selects and commits trajectories." claim_type: architectural_commitment subject: E3.trajectory_commitment polarity: asserts status: active lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-25T16:56:17.901452Z adjudication_decision_status: applied adjudication_recommendation: retain_ree depends_on: - INV-012 - ARC-001 - ARC-002 - ARC-004 - ARC-005 location: docs/architecture/e3.md#arc-003 source: - docs/processed/legacy_tree/docs/architecture/e3.md - docs/processed/legacy_tree/architecture/E3.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-24_prefrontal_primitives.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: ARC-004 title: "L-space is a multi-timescale latent stack." claim_type: architectural_commitment subject: L_space.latent_stack polarity: asserts status: active depends_on: - INV-013 - INV-002 location: docs/architecture/l_space.md#arc-004 source: - docs/processed/legacy_tree/docs/architecture/l_space.md - docs/processed/legacy_tree/architecture/latent_stack.md - id: ARC-005 title: "Control plane routes precision and modes." claim_type: architectural_commitment subject: control_plane.precision_routing polarity: asserts status: active depends_on: - INV-008 - INV-009 - INV-014 - ARC-004 location: docs/architecture/control_plane.md#arc-005 source: - docs/processed/legacy_tree/docs/architecture/control_plane.md - docs/processed/legacy_tree/architecture/control_plane.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-09_empathy.md - id: ARC-006 title: "Entities are sparse, persistent, bindable structures." claim_type: architectural_commitment subject: entities.binding polarity: asserts status: provisional depends_on: - ARC-004 - INV-002 location: docs/architecture/entities_and_binding.md#arc-006 source: - docs/processed/legacy_tree/docs/architecture/entities_and_binding.md - docs/processed/legacy_tree/DANIEL_README.md - id: ARC-007 title: "Hippocampal systems store and replay paths through residue-field terrain." claim_type: architectural_commitment subject: hippocampus.path_memory polarity: asserts status: active lifecycle_stage: adjudicated adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: hybridize evidence_quality_note: | V2 EXQ (path_memory_ablation, 2026-03-08): FAIL (partial_support=true, 1/2 criteria). PATH_MEMORY vs PATH_ABLATED agent harm: 0.896 vs 0.912 -- sub-threshold difference. Proxy representation was a 26-element obs-vector slice [252, 278] standing in for hippocampal path memory. This is too crude to test the architectural claim: proper ARC-007 path memory requires a functioning HippocampalModule navigating a structured path space (SD-004). The ablation test cannot distinguish "path memory doesn't help" from "the proxy is too weak to carry meaningful path information." V3 gate cleared (2026-03-28): SD-004 action objects + HippocampalModule now implemented in ree-v3. V3 path-memory experiments can proceed. EXQ-114 PASS (2026-03-28, 4/4): MAP_NAV harm_rate=0.006 vs MAP_ABLATED=0.765. 99.2% harm reduction. 2/2 seeds consistent. Strong V3 confirmation of ARC-007. Q-020 resolved (2026-04-02): The "no value computation" constraint is compatible with hippocampal geometry being value-shaped. The constraint specifies only that the computation of value (RPE, utility) occurs in BG/amygdala/PFC, not in hippocampus. Hippocampus stores the geometrically encoded result via BTSP-mediated write operations (Bittner 2017, Science). MECH-073 and ARC-007 are co-true under Resolution A. Literature: Dupret 2010, Gauthier 2018, Bittner 2017, Teyler & Rudy 2007. depends_on: - ARC-004 - ARC-003 - ARC-018 location: docs/architecture/hippocampal_systems.md#arc-007 source: - docs/processed/legacy_tree/docs/architecture/hippocampal_braid.md - docs/processed/legacy_tree/architecture/Hippocampal_braid.md - docs/processed/legacy_tree/docs/architecture/hippocampal_systems.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - id: ARC-008 title: "Commitment eligibility is gated by tau, rho, and phi." claim_type: architectural_commitment subject: temporal.depth_phase_gating polarity: asserts status: provisional depends_on: - INV-002 - ARC-004 - ARC-005 - ARC-003 location: docs/architecture/temporal_dynamics.md#arc-008 source: - docs/processed/legacy_tree/architecture/depth_phase_spec.md - docs/processed/legacy_tree/docs/architecture/temporal_dynamics.md - id: IMPL-001 title: "Glossary of canonical REE terms." claim_type: implementation_note subject: repo.glossary polarity: asserts status: active depends_on: [] location: docs/glossary.md#impl-001 source: - docs/processed/legacy_tree/docs/glossary.md - id: Q-001 title: "What mechanisms produce entity emergence and binding?" claim_type: open_question subject: entities.emergence_mechanism polarity: asserts status: active depends_on: - ARC-006 location: docs/architecture/entities_and_binding.md#q-001 source: - docs/processed/legacy_tree/docs/architecture/entities_and_binding.md evidence_quality_note: | EXQ-146 FAIL/weakens (2026-03-29): entity binding discriminative pair. First experimental entry. Failure does not yet constrain the question -- substrate and training budget gaps likely. - id: ARC-009 title: "Language is a symbolic mediation and coordination layer." claim_type: architectural_commitment subject: language.symbolic_mediation polarity: asserts status: active depends_on: - INV-003 - INV-007 - ARC-010 - ARC-005 - ARC-004 location: docs/architecture/language.md#arc-009 source: - docs/processed/legacy_tree/architecture/language/README.md - docs/processed/legacy_tree/architecture/language/language_contract.md - docs/processed/legacy_tree/architecture/language/language_functions.md - docs/processed/legacy_tree/docs/architecture/language.md - docs/thoughts/2026-02-09_language.md - docs/thoughts/2026-02-09_arcuate_fasciculus_language_nudges.md - id: ARC-010 title: "Social cognition uses mirror modelling and coupling." claim_type: architectural_commitment subject: social.mirror_modelling polarity: asserts status: active depends_on: - INV-005 - ARC-004 - ARC-006 location: docs/architecture/social.md#arc-010 source: - docs/processed/legacy_tree/architecture/social/README.md - docs/processed/legacy_tree/architecture/social/mirror_modelling.md - docs/processed/legacy_tree/architecture/social/social_coupling.md - docs/processed/legacy_tree/architecture/social/otherness_inference.md - docs/processed/legacy_tree/docs/architecture/social.md - docs/thoughts/2026-02-09_empathy.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: ARC-011 title: "Offline integration (sleep) is required for stability." claim_type: architectural_commitment subject: sleep.offline_integration polarity: asserts status: active depends_on: - INV-010 - ARC-007 - ARC-005 location: docs/architecture/sleep.md#arc-011 source: - docs/processed/legacy_tree/architecture/sleep/README.md - docs/processed/legacy_tree/architecture/sleep/sleep_contract.md - docs/processed/legacy_tree/docs/architecture/sleep.md - id: ARC-012 title: "E3 does not require an explicit ethical cost term." claim_type: architectural_commitment subject: ethics.no_explicit_cost_term polarity: asserts status: active depends_on: - INV-001 - ARC-003 - ARC-005 - ARC-007 location: docs/architecture/e3.md#arc-012 source: - docs/processed/legacy_tree/REE_CORE.md - id: MECH-001 title: "Astrocyte-aware regulatory stack mediates control-plane precision routing." claim_type: mechanism_hypothesis subject: control_plane.astrocyte_regulatory_stack polarity: asserts status: candidate depends_on: - ARC-005 - INV-008 - ARC-004 location: docs/architecture/astrocyte_regulatory_stack.md#mech-001 source: - docs/processed/legacy_tree/docs/astrocyte_aware_regulatory_stack/regulatory_stack_model.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/regulatory_stack_model.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/README.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/astrocytes_in_brief.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/implementation_hooks.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/ree_component_mapping.md - docs/processed/legacy_tree/architecture/astrocyte_aware_regulatory_stack/references.md - docs/processed/legacy_tree/docs/architecture/astrocyte_regulatory_stack.md - id: Q-002 title: "What is the appropriate spatial resolution for R(x,t)?" claim_type: open_question subject: astrocyte.R_field_resolution polarity: asserts status: active depends_on: - MECH-001 location: docs/architecture/astrocyte_regulatory_stack.md#q-002 source: - docs/processed/legacy_tree/docs/astrocyte_aware_regulatory_stack/regulatory_stack_model.md evidence_quality_note: | EXQ-147 FAIL/weakens via MECH-128 (2026-03-29): E1 goal conditioning pair FAIL. First experimental entry. Failure likely reflects training budget / substrate gap rather than fundamental resolution question. EXQ-170 FAIL 1/3 x2 replicated (2026-03-30, two independent runs): fine resolution (32 RBF centers) vs coarse (4 centers). Fine harm_rate=0.028 vs coarse 0.0003 (93x worse). Fine residue_accuracy=0.016 (threshold 0.10). Consistent across both runs. z_world latent lacks spatial organization to support fine-grained RBF accumulation at V3 scale -- 32 centers spread learning signal too thinly; 4 centers sufficient for zone-level harm/safe distinction. V3 scale finding: coarse resolution adequate, fine counterproductive. Does not rule out fine resolution at higher task complexity or after SD-005/higher-dim z_world. Evidence direction: weakens (fine-resolution hypothesis at V3 scale). EXQ-215 FAIL 2/5 (2026-04-03): Residue resolution discriminative pair FAIL. HIGH_RES (64 RBF centers) vs LOW_RES (8 centers). C1 FAIL (per-seed harm direction inconsistent across seeds). C2 PASS (mean_harm_delta=0.056: high_res_harm=0.033 vs low_res_harm=0.090 -- high-res has LOWER harm, C2 passes threshold). C3 FAIL (residue_accuracy: high_res=-0.010 vs low_res=0.033 -- low-res BETTER at hazard correlation). C4/C5 PASS. Surprising: high-res outperforms on harm avoidance (C2) but underperforms on residue accuracy (C3 reverse). Confirms Q-002 complexity: fine resolution improves harm avoidance but degrades residue-hazard mapping quality. The question remains open but evidence tilts toward: fine resolution better for harm avoidance behavior, coarse better for terrain map quality. Consistent with EXQ-170 finding (coarse adequate). Evidence direction: weakens fine-resolution-is-universally-better hypothesis. - id: Q-003 title: "Should R(x,t) be scalar or vector?" claim_type: open_question subject: astrocyte.R_field_dimensionality polarity: asserts status: active depends_on: - MECH-001 location: docs/architecture/astrocyte_regulatory_stack.md#q-003 source: - docs/processed/legacy_tree/docs/astrocyte_aware_regulatory_stack/regulatory_stack_model.md evidence_quality_note: | EXQ-148 FAIL/weakens x2 (2026-03-28, 2026-03-29): R-field dimensionality discriminative pair run twice, both FAIL/weakens. Two consistent results. Weakening evidence accumulating; question may need reframing once residue field contrast infrastructure is confirmed working. - id: Q-004 title: "How to calibrate tau_R relative to E1/E2?" claim_type: open_question subject: astrocyte.tau_R_calibration polarity: asserts status: active depends_on: - MECH-001 location: docs/architecture/astrocyte_regulatory_stack.md#q-004 source: - docs/processed/legacy_tree/docs/astrocyte_aware_regulatory_stack/regulatory_stack_model.md evidence_quality_note: | EXQ-149 PARTIAL (2026-03-29): tau_R in [0.02, 0.5] does not affect harm rate at current training budget. FAST_TAU produces ~400x more residue contrast than SLOW_TAU (residue_contrast=3.09 vs 0.008), but all conditions show identical mean_harm (~9.085). Question shifts: tau_R matters for residue accumulation patterns and timescales, not harm rate at this training scale. Calibration question remains open at higher training budgets. - id: Q-005 title: "Can sleep anneal or reset R(x,t)?" claim_type: open_question subject: astrocyte.sleep_annealing polarity: asserts status: active depends_on: - MECH-001 location: docs/architecture/astrocyte_regulatory_stack.md#q-005 source: - docs/processed/legacy_tree/docs/astrocyte_aware_regulatory_stack/regulatory_stack_model.md evidence_quality_note: | EXQ-150 FAIL/mixed (2026-03-29): superseded -- sleep not implemented in V3; post-training phase simulation (anneal vs no-anneal) not representative of genuine sleep annealing. Manifest marked superseded. Anneal vs no-anneal conditions showed identical harm (9.050). Untestable until V3 sleep substrate implemented. - id: ARC-013 title: "Residue is persistent latent-space curvature; hippocampal paths form a cognitive map." claim_type: architectural_commitment subject: residue.geometry polarity: asserts status: active depends_on: - INV-006 - INV-004 - ARC-004 location: docs/architecture/residue_geometry.md#arc-013 source: - docs/processed/legacy_tree/docs/architecture/residue_geometry.md - docs/processed/legacy_tree/architecture/residue_geometry.md - docs/thoughts/2026-02-08_residue_paths_cognitive_map.md - id: ARC-014 title: "Default Mode enables safe imagination without commitment." claim_type: architectural_commitment subject: default_mode.internal_generative polarity: asserts status: active depends_on: - INV-011 - ARC-003 - ARC-005 - ARC-007 - ARC-013 location: docs/architecture/default_mode.md#arc-014 source: - docs/processed/legacy_tree/docs/architecture/default_mode.md - docs/processed/legacy_tree/architecture/Default_mode.md - id: IMPL-002 title: "Repository metadata and contribution process." claim_type: implementation_note subject: repo.meta polarity: asserts status: active depends_on: [] location: docs/repo_meta.md#impl-002 source: - docs/processed/legacy_tree/CITATION.cff - docs/processed/legacy_tree/CONTRIBUTING.md - docs/processed/legacy_tree/.github/ISSUE_TEMPLATE/feature_request.yml - docs/processed/legacy_tree/.github/ISSUE_TEMPLATE/bug_report.yml - docs/processed/legacy_tree/.github/ISSUE_TEMPLATE/discussion.yml - docs/processed/legacy_tree/src/placeholder.md - id: IMPL-003 title: "Minimum instantiation specification." claim_type: implementation_note subject: spec.minimum polarity: asserts status: active depends_on: - ARC-001 - ARC-002 - ARC-003 - ARC-004 - ARC-005 - ARC-013 - ARC-011 location: docs/REE_MIN_SPEC.md#impl-003 source: - docs/processed/legacy_tree/docs/REE_MIN_SPEC.md - id: IMPL-004 title: "Legacy REE overview summary." claim_type: implementation_note subject: overview.legacy polarity: asserts status: legacy depends_on: - INV-001 - INV-002 - INV-003 - ARC-001 - ARC-002 - ARC-003 - ARC-004 - ARC-007 - ARC-018 location: docs/REE_overview.md#impl-004 source: - docs/processed/legacy_tree/docs/REE_overview.md - id: IMPL-005 title: "Failure mode taxonomy." claim_type: implementation_note subject: failure_modes polarity: asserts status: active depends_on: - INV-006 - INV-008 - ARC-005 - ARC-013 - ARC-010 - ARC-007 - ARC-018 location: docs/REE_failure_modes.md#impl-005 source: - docs/processed/legacy_tree/docs/REE_failure_modes.md - docs/thoughts/FAILURE-2026-02-12_COORDINATE-SYSTEM-FOR-COGNITIVE-PATHOLOGY.md - docs/thoughts/2026-02-12_DEPRESSIVE-PATH-PRUNING-HIPPOCAMPAL-ROLLBACK.md - id: IMPL-006 title: "Legacy migration mapping." claim_type: implementation_note subject: migration.legacy polarity: asserts status: legacy depends_on: [] location: docs/MIGRATION.md#impl-006 source: - docs/processed/legacy_tree/docs/MIGRATION.md - id: IMPL-007 title: "Legacy refactor final output summary." claim_type: implementation_note subject: refactor.final_output.legacy polarity: asserts status: legacy depends_on: [] location: docs/FINAL_OUTPUT.md#impl-007 source: - docs/processed/legacy_tree/docs/FINAL_OUTPUT.md - id: IMPL-008 title: "Program phases, repository roles, and phase-gate criteria." claim_type: implementation_note subject: roadmap.program_phases polarity: records status: candidate depends_on: [] location: docs/roadmap.md#impl-008 source: - docs/roadmap.md - docs/processed/legacy_tree/roadmap.md - id: IMPL-009 title: "Wiring notes and cross-reference summary." claim_type: implementation_note subject: wiring.notes.legacy polarity: asserts status: legacy depends_on: [] location: docs/notes/wiring_notes.md#impl-009 source: - docs/processed/legacy_tree/WIRING_NOTES.md - id: IMPL-010 title: "Android world environment contract." claim_type: implementation_note subject: examples.android_world polarity: asserts status: active depends_on: - IMPL-003 location: docs/examples/android_world_environment.md#impl-010 source: - docs/processed/legacy_tree/examples/android_world/environment.md - id: IMPL-011 title: "Toy world environment contract." claim_type: implementation_note subject: examples.toy_world polarity: asserts status: active depends_on: - IMPL-003 location: docs/examples/toy_world_environment.md#impl-011 source: - docs/processed/legacy_tree/examples/toy_world/environment.md - id: IMPL-012 title: "Toy world scoring functions." claim_type: implementation_note subject: examples.toy_world_scoring polarity: asserts status: active depends_on: - IMPL-003 location: docs/examples/toy_world_scoring_functions.md#impl-012 source: - docs/processed/legacy_tree/examples/toy_world/scoring_functions.md - id: IMPL-013 title: "Documentation operating procedure and prompts." claim_type: implementation_note subject: docs.readme polarity: asserts status: active depends_on: [] location: docs/README.md#impl-013 source: - docs/processed/legacy_tree/docs/README.md - id: IMPL-014 title: "Documentation change history." claim_type: implementation_note subject: docs.changelog polarity: asserts status: active depends_on: [] location: docs/changelog.md#impl-014 source: - docs/processed/legacy_tree/docs/changelog.md - id: IMPL-015 title: "Legacy architecture overview." claim_type: implementation_note subject: architecture.overview.legacy polarity: asserts status: legacy depends_on: [] location: docs/architecture/overview.md#impl-015 source: - docs/processed/legacy_tree/architecture/README_architecture.md - id: IMPL-016 title: "Trajectory selection detail for E3." claim_type: implementation_note subject: trajectory.selection.detail polarity: asserts status: candidate depends_on: - ARC-003 - ARC-004 - ARC-013 location: docs/architecture/trajectory_selection.md#impl-016 source: - docs/processed/legacy_tree/architecture/trajectory_selection.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - id: MECH-002 title: "Precision control analogues shape cognitive regimes." claim_type: mechanism_hypothesis subject: precision.control_analogues polarity: asserts status: provisional depends_on: - ARC-005 - INV-008 - ARC-004 location: docs/architecture/precision_control.md#mech-002 source: - docs/processed/legacy_tree/architecture/precision_control.md - id: MECH-003 title: "Precision must be tau-scoped with lossy projections." claim_type: mechanism_hypothesis subject: precision.tau_scoping polarity: asserts status: provisional depends_on: - INV-008 - ARC-004 location: docs/architecture/precision_scoping.md#mech-003 source: - docs/processed/legacy_tree/architecture/precision_scoping.md - id: MECH-004 title: "Signal-to-knob wiring map for control plane." claim_type: mechanism_hypothesis subject: control_plane.signal_map polarity: asserts status: candidate depends_on: - ARC-005 - ARC-003 - ARC-017 - MECH-037 - MECH-062 location: docs/architecture/control_plane_signal_map.md#mech-004 source: - docs/processed/legacy_tree/architecture/control_plane_signal_map.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-17_control_plane_update.md - docs/thoughts/17-02-26_necessary_separations_based_on_considering-prompt_injection.md - id: MECH-005 title: "Path authority and interruptibility via norepinephrine-like control." claim_type: mechanism_hypothesis subject: control_plane.path_authority polarity: asserts status: provisional depends_on: - ARC-005 - ARC-003 - ARC-008 location: docs/architecture/path_authority_and_interrupts.md#mech-005 source: - docs/processed/legacy_tree/architecture/path_authority_and_interrupts.md - id: MECH-006 title: "Serotonin-like modulation governs representational collapse." claim_type: mechanism_hypothesis subject: control_plane.serotonin_collapse polarity: asserts status: provisional depends_on: - ARC-005 - ARC-008 location: docs/architecture/serotonin.md#mech-006 source: - docs/processed/legacy_tree/architecture/serotonin.md - id: MECH-007 title: "Attention must be fragmented across control axes." claim_type: mechanism_hypothesis subject: attention.fragmentation polarity: asserts status: provisional depends_on: - INV-009 - ARC-005 location: docs/architecture/why_attention_must_be_fragmented.md#mech-007 source: - docs/processed/legacy_tree/architecture/why_attention_must_be_fragmented.md - id: MECH-008 title: "Legacy mechanism: superseded by AA/PCM and control-channel mode framing." claim_type: mechanism_hypothesis subject: control_plane.mode_management polarity: asserts status: legacy depends_on: - ARC-005 - ARC-003 location: docs/architecture/mode_manager.md#mech-008 source: - docs/processed/legacy_tree/mode_manager.md - id: MECH-010 title: "Language emergence and bootstrapping dynamics." claim_type: mechanism_hypothesis subject: language.emergence_bootstrapping polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/emergence_and_bootstrapping.md#mech-010 source: - docs/processed/legacy_tree/architecture/language/emergence_and_bootstrapping.md - docs/thoughts/2026-02-09_language.md - id: MECH-011 title: "Language and learning dynamics." claim_type: mechanism_hypothesis subject: language.and_learning polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/language_and_learning.md#mech-011 source: - docs/processed/legacy_tree/architecture/language/language_and_learning.md - id: MECH-012 title: "Language and institutions interplay." claim_type: mechanism_hypothesis subject: language.and_institutions polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/language_and_institutions.md#mech-012 source: - docs/processed/legacy_tree/architecture/language/language_and_institutions.md - id: MECH-013 title: "Language failure modes and pathologies." claim_type: mechanism_hypothesis subject: language.failure_modes polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/language_failure_modes.md#mech-013 source: - docs/processed/legacy_tree/architecture/language/language_failure_modes.md - id: MECH-014 title: "Minimal signalling channel requirements." claim_type: mechanism_hypothesis subject: language.minimal_signalling_channel polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/minimal_signalling_channel.md#mech-014 source: - docs/processed/legacy_tree/architecture/language/minimal_signalling_channel.md - id: MECH-015 title: "Trust and deception dynamics." claim_type: mechanism_hypothesis subject: language.trust_and_deception polarity: asserts status: candidate depends_on: - ARC-009 - ARC-010 location: docs/architecture/language/trust_and_deception.md#mech-015 source: - docs/processed/legacy_tree/architecture/language/trust_and_deception.md - id: MECH-016 title: "Precision recalibration during sleep." claim_type: mechanism_hypothesis subject: sleep.precision_recalibration polarity: asserts status: candidate depends_on: - ARC-011 - ARC-007 location: docs/architecture/sleep/precision_recalibration.md#mech-016 source: - docs/processed/legacy_tree/architecture/sleep/precision_recalibration.md - id: MECH-017 title: "Reality consolidation during sleep." claim_type: mechanism_hypothesis subject: sleep.reality_consolidation polarity: asserts status: candidate depends_on: - ARC-011 - ARC-007 location: docs/architecture/sleep/reality_consolidation.md#mech-017 source: - docs/processed/legacy_tree/architecture/sleep/reality_consolidation.md - id: MECH-018 title: "Residue integration during sleep." claim_type: mechanism_hypothesis subject: sleep.residue_integration polarity: asserts status: candidate depends_on: - ARC-011 - ARC-007 location: docs/architecture/sleep/residue_integration.md#mech-018 source: - docs/processed/legacy_tree/architecture/sleep/residue_integration.md - id: IMPL-017 title: "Conflict index and resolution entry point." claim_type: implementation_note subject: conflicts.index polarity: asserts status: active depends_on: - Q-008 - Q-009 - Q-011 location: docs/conflicts/README.md#impl-017 source: - docs/processed/legacy_tree/docs/conflicts/ethics_module_vs_cost_term.md - docs/conflicts/valence_vectors_vs_mu_kappa_overlays.md - docs/conflicts/care_override_vs_other_harm_veto.md - docs/conflicts/rollout_entropy_floor_vs_residue_persistence.md - docs/conflicts/resolutions/2026-02-08_ethics-module-vs-cost-term.md - id: IMPL-018 title: "Claim index and navigation." claim_type: implementation_note subject: claims.index polarity: asserts status: active depends_on: [] location: docs/claims/claim_index.md#impl-018 source: - docs/processed/legacy_tree/docs/claims/claim_index.md - id: IMPL-019 title: "Self-first, social-later developmental testing order heuristic." claim_type: implementation_note subject: developmental_ordering.self_first_social_later polarity: asserts status: candidate depends_on: - ARC-019 - Q-006 - ARC-005 - ARC-007 - ARC-010 location: docs/architecture/developmental_curriculum.md#impl-019 source: - docs/thoughts/DEV-ROADMAP-SELF-FIRST-SOCIAL-LATER.md - id: INV-018 title: "Agency is required; passive predictors are not REE." claim_type: invariant subject: agency.requirement polarity: asserts status: stable depends_on: [] location: docs/invariants.md#inv-018 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: ARC-015 title: "Self-impact attribution and responsibility flow are required." claim_type: architectural_commitment subject: agency.self_impact_attribution polarity: asserts status: provisional depends_on: - INV-018 - INV-012 - ARC-003 - ARC-005 - ARC-004 - ARC-013 - ARC-007 location: docs/architecture/agency_responsibility_flow.md#arc-015 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-019 title: "Control plane shapes modes of cognition, not discrete choices." claim_type: mechanism_hypothesis subject: control_plane.mode_shaping polarity: asserts status: provisional depends_on: - ARC-005 location: docs/architecture/control_plane.md#mech-019 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-020 title: "Legacy mechanism: superseded by control-plane mode framing." claim_type: mechanism_hypothesis subject: cognitive_modes.emergent polarity: asserts status: legacy depends_on: - ARC-005 location: docs/architecture/mode_manager.md#mech-020 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-021 title: "Subjective now is a control surface across temporal horizons." claim_type: mechanism_hypothesis subject: temporal.now_construction polarity: asserts status: provisional depends_on: - ARC-008 - ARC-005 location: docs/architecture/temporal_dynamics.md#mech-021 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-022 title: "Hippocampal systems inject hypotheses gated by control plane." claim_type: mechanism_hypothesis subject: hippocampal.hypothesis_injection polarity: asserts status: provisional depends_on: - ARC-007 - ARC-005 location: docs/architecture/hippocampal_systems.md#mech-022 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-023 title: "Responsibility is geometric and path-dependent." claim_type: mechanism_hypothesis subject: responsibility.geometry polarity: asserts status: provisional depends_on: - ARC-013 - ARC-015 location: docs/architecture/agency_responsibility_flow.md#mech-023 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: MECH-024 title: "Selfhood, personality, and ethics converge structurally." claim_type: mechanism_hypothesis subject: selfhood_personality_ethics.convergence polarity: asserts status: provisional depends_on: - ARC-015 location: docs/architecture/agency_responsibility_flow.md#mech-024 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md - id: Q-006 title: "Is ethics developmental rather than additive?" claim_type: open_question subject: ethics.developmental polarity: asserts status: active depends_on: - ARC-015 location: docs/architecture/agency_responsibility_flow.md#q-006 source: - docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md evidence_quality_note: | EXQ-151 FAIL/mixed (2026-03-29): implementation gap -- residue field not building contrast (guidance="implementation_problem_residue_field_not_building"); all conditions delta_harm=0; not informative about Q-006 itself. Scoring excluded. Infrastructure fix required: residue field must build meaningful contrast before developmental vs additive ethics can be tested. - id: Q-007 title: "Do affect dimensions (not necessarily universal expressions) map to stable control-channel regimes in z_beta?" claim_type: open_question subject: emotion.affect_dimension_channel_mapping polarity: asserts status: active evidence_quality_note: | V3 gate cleared (2026-03-28): SD-005 (z_self/z_world split) and ARC-016 (dynamic precision) now implemented in ree-v3. V3 experiments can now test z_beta/running_variance correlation. V2 experiment EXQ-024 (valence_regime_correlation, EVB-0006) FAIL -- run twice (2026-03-08, 2026-03-15), consistent result. Proximate failure: E3 precision hardcoded at 0.5 (no dynamic precision channel implemented), making the correlation mathematically impossible to observe. Fundamental blocker: V2 lacks correct E1/E2/E3 representation — z_beta receives mixed signals from conflated learning channels. Fixing precision alone would not produce valid results. Proxy problem: experiment used health/energy (external environment state) as valence signal; correct test requires z_beta valence dimension (internal affective representation) correlated with E3 precision. No valid evidence in current corpus. EXQ-051 FAIL (2026-03-20): Pearson r(z_beta, running_variance) = -0.033 (volatile), -0.029 (stable). Root cause confirmed — architectural gap: z_beta is computed as beta_encoder(cat(z_self, z_world)) with no pathway from E3's running_variance. Volatile environments change hazard positions but not sensory statistics (z_beta_norm ≈ 0.72 in both conditions). Not a threshold or calibration issue. Literature pull (2026-03-20): gap is well-grounded in neuromodulatory literature. Yu & Dayan 2005 (PMID 15944135, Neuron): LC-NE encodes unexpected uncertainty — running mismatch between predicted and actual variance. Not a sensory-content signal; NE rises when the harm-prediction distribution itself has changed. Running_variance in E3 is the direct REE computational analog. Payzan-LeNormand et al. 2013 (PMID 24312028): LC-NE drives exploration and prior resetting under high prediction-error variance. Mathys HGF model: affective state maps to log-volatility estimate (μ₃, second level), not first-level prediction error. Barrett/Seth interoceptive inference: arousal = magnitude of interoceptive prediction error, NOT exteroceptive sensory content. Fix roadmap (2026-03-20): Option A (EXQ-051b, immediate): LatentStackConfig.volatility_signal_dim=1 injects E3._running_variance as scalar into beta_encoder. Tests whether z_beta shows condition difference and Pearson r(rv, z_beta) > 0.3. Option C (post SD-010, long-term): z_beta driven from harm_history — harm stream as REE's interoceptive/allostatic channel (Barrett arousal = |interoceptive PE|). running_harm_variance becomes μ₃ analog. Requires SD-010 first. EXQ-051b queued 2026-03-20 (V3-EXQ-051b). EXQ-051c FAIL 1/3 (2026-03-30, V3-EXQ-051c): C1 PASS -- volatile z_beta_norm=0.775 vs stable=0.659, delta=+0.116 (threshold >0.05 met). C2 FAIL -- r(z_beta, mean_rv)=0.000; mean_rv is flat at 0.500 in both conditions (metric broken -- rv not varying with env). C3 FAIL -- consequence of C2 failure. Directional signal confirmed for z_beta condition difference (C1), but rv injection pathway not producing variance signal. v3_pending hold stands until rv metric is repaired and C2/C3 retested. REFRAME (2026-04-02): Literature pull surfaced Barrett et al. 2019 (weakens universal expression premise) and Barrett 2017 (constructed emotion theory, mixed). Facial configurations are not reliably diagnostic of emotional states across cultures. The constructionist position argues emotions are constructed categories, not natural kinds with fixed neural signatures. This challenges the "universal expressions" framing but NOT the core architectural question of whether z_beta regimes exist. Reframed from "universal expressions -> regimes" to "affect dimensions (valence/arousal) -> regimes." Dimensional affect (Russell 1980 circumplex, Barrett dimensional models) is more consistent with REE's z_beta architecture than categorical emotion. z_beta channels may represent valence and arousal dimensions rather than discrete emotional categories. This reframe preserves the testable core (stable z_beta regime structure) while dropping the weakened universal-expression assumption. depends_on: - ARC-005 - ARC-021 - MECH-039 - MECH-040 - MECH-069 location: docs/architecture/control_plane.md#q-007 source: - docs/architecture/control_plane.md - id: Q-008 title: "Legacy: resolved in favor of valence/mu-kappa separation with calibration follow-up." claim_type: open_question subject: valence_vs_mu_kappa_overlays polarity: asserts status: legacy depends_on: - ARC-017 - MECH-035 - MECH-048 location: docs/architecture/control_plane.md#q-008 source: - docs/architecture/control_plane.md - id: Q-009 title: "Legacy: resolved via bounded care-override policy with hard catastrophic other-harm veto." claim_type: open_question subject: care_weights_override_other_harm_veto polarity: asserts status: legacy depends_on: - MECH-036 - MECH-052 - ARC-010 location: docs/architecture/social.md#q-009 source: - docs/architecture/social.md - id: Q-010 title: "Legacy: separation question resolved into MECH-055." claim_type: open_question subject: hedonic_valence_signed_pe_separation polarity: asserts status: legacy depends_on: - MECH-035 - MECH-048 - MECH-054 location: docs/architecture/control_plane.md#q-010 source: - docs/architecture/control_plane.md - id: Q-011 title: "Legacy: resolved by placing diversity-floor control in pre-commit/offline sampling." claim_type: open_question subject: hippocampus.minimum_rollout_entropy_floor polarity: asserts status: legacy depends_on: - ARC-018 - ARC-005 - ARC-011 location: docs/architecture/hippocampal_systems.md#q-011 source: - docs/architecture/hippocampal_systems.md - docs/thoughts/2026-02-12_DEPRESSIVE-PATH-PRUNING-HIPPOCAMPAL-ROLLBACK.md - id: ARC-016 title: "Modes are control-plane regimes applied to shared predictive machinery: the precision-to-commitment circuit." claim_type: architectural_commitment subject: cognitive_modes.control_plane_regimes polarity: asserts status: provisional status_note: | Split 2026-03-22: ARC-016 now covers only the structural/mechanistic circuit (E3-derived variance -> relative threshold -> BetaGate -> action_selection). The behavioral consequence layer (committed vs uncommitted -> measurably distinct harm outcomes) is ARC-029. Provisional status is justified by EXQ-018b PASS 5/5 and EXQ-060 PASS 4/5. evidence_quality_note: | Scope (post-split 2026-03-22): ARC-016 asserts that operating modes exist as control-plane regimes — specifically that E3-derived prediction variance drives a commitment threshold which gates action-selection via BetaGate. The behavioral consequence of that gating (harm differences between modes) is ARC-029. V2 EXQ (precision_regime_probe, 2026-03-08): FAIL (partial_support=true, 1/2 criteria). Criterion 1 met: precision gap achieved (1.697, easily). Criterion 2 failed: behavioral distinction absent. Root cause: precision was externally imposed (not E3-derived) and commitment-to-action circuit not wired end-to-end in V2. Two V3 requirements exposed: (1) E3-derived dynamic precision calibrated from prediction error; (2) commitment gating must actually change action-selection. Note: MECH-059 (precision/PE structural separation) PASSED in V2, confirming the channels exist as distinct signals. ARC-016 structural separation is necessary but not sufficient for ARC-029 behavioral distinction. V3 EXQ-018 FAIL (2026-03-20): Variance mechanism working but absolute threshold (0.40) was 100x the operating variance range. Fix: relative threshold. V3 EXQ-018b PASS (2026-03-20, 5/5 criteria): Relative threshold confirmed working. commit_threshold = 2 x training_baseline_variance. Stable env: commit_rate=0.90, precision=718; perturbed env: commit_rate=0.50, precision=426. 40% precision drop produces proportional 40-point commit-rate drop. Circuit is end-to-end: E3-derived variance -> commit_threshold comparison -> BetaGate -> action_selection. EXQ-060 PASS (2026-03-21, 4/5): Committed-condition BetaGate confirmed. committed_step_count=5980, hold_rate_during_committed=0.936, calibration_gap_approach=0.930. C4 missed (propagation_count=0) because single-agent design keeps agent permanently committed — not a mechanism failure. ARC-016 core circuit is validated. See ARC-029 for the behavioral consequence layer. depends_on: - ARC-005 - MECH-059 - ARC-021 location: docs/architecture/modes_of_cognition.md#arc-016 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: ARC-017 title: "Minimal stream tags with typed exteroception and explicit reality-coherence lane." claim_type: architectural_commitment subject: sensory_stream_tags.minimal_set polarity: asserts status: provisional evidence_quality_note: | EXQ-129 FAIL (2026-03-29): stream tag pair FAIL. First experimental entry. EXQ-135 FAIL (2026-03-29): reality coherence pair FAIL. Second FAIL on this claim. depends_on: - INV-008 - INV-012 - ARC-004 - ARC-005 - ARC-003 - ARC-015 location: docs/architecture/sensory_stream_tags.md#arc-017 source: - docs/thoughts/2026-02-08_sensory_stream_tags.md - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-09_empathy.md - docs/thoughts/2026-02-17_control_plane_update.md - docs/thoughts/17-02-26_necessary_separations_based_on_considering-prompt_injection.md - id: ARC-018 title: "Hippocampus generates explicit rollouts and post-commitment viability mapping." claim_type: architectural_commitment subject: hippocampus.rollout_viability_mapping polarity: asserts status: active lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree v3_pending: false reframe_note: | Original framing (2026-02-09) incorrectly attributed viability mapping to E1 prediction error (SELF_SENSORY mismatch). V2 EXQ-021 FAIL (2026-03-08, 2026-03-15) confirmed this framing is wrong: E1 prediction error has no advantage over a frozen E1 for harm navigation. Corrected framing (2026-03-15): hippocampus builds viability map indexed by E2 action-object coordinates, updated by E3 harm/goal error. E1 provides perceptual context to E2 only. E2's latent space is action consequences (action objects), not sensory state transitions. The map lives in hippocampal geometry, not in E1. evidence_quality_note: | V2 experiment EXQ-021 (rollout_viability_mapping, EVB-0017) FAIL — run twice (2026-03-08, 2026-03-15), consistent result. VIABILITY_MAPPED (E1 updating) shows no advantage over VIABILITY_FIXED (frozen E1) in harm navigation; FIXED outperforms MAPPED in 2/3 seeds. This is a conceptual failure of the original claim framing, not a substrate or tuning issue. V3 EXQ-042 PASS (2026-03-19): terrain_prior learns E3 preferences, hippo_quality_gap=0.393, terrain_loss→0. V3 EXQ-053 PASS (2026-03-20): terrain-guided navigation produces 50× less harm than random (harm_per_step 0.00147 vs 0.0724, viability_advantage=0.071, 5/5 criteria). SD-004 and SD-005 prerequisites implemented in V3 substrate. v3_pending cleared 2026-03-20. Outstanding: EXQ-046 (ARC-007 ablation) will confirm the action-object indexing mechanism specifically contributes. EXQ-120 PASS (2026-03-28): viability map discriminative pair. evidence_direction=supports. Adds one more support entry confirming the V3 substrate continues to pass this claim. EXQ-172 INCONCLUSIVE (2026-03-30, T213638Z): E2 quality borderline (r2=0.203, gate=0.15); direction criteria fail. Not a clean test. Superseded by T070425Z run. EXQ-172 FAIL 0/4 (2026-03-30, T070425Z): harm_advantage_mean=-0.190. E2 quality adequate (r2=0.203) -- rollout planning counterproductive at current fidelity. ARC-018 may be E2-quality-gated. Evidence direction: weakens. EXQ-196 FAIL (2026-04-04): harm_advantage_mean=0.0 across all 3 seeds with good substrate quality (e2_world_r2=0.766, residue populated). Classified non-contributory (2026-04-06 governance): zero advantage is diagnostic of missing SD-004 substrate (action-object hippocampal map backbone), not refutation of ARC-018. Verdict deferred pending SD-004. depends_on: - ARC-007 - ARC-003 - ARC-002 - ARC-001 - ARC-021 location: docs/architecture/hippocampal_systems.md#arc-018 source: - docs/thoughts/2026-02-09_starting_with_sensory_streams.md - docs/thoughts/2026-02-09_e2_hpc_interface.md - docs/thoughts/2026-03-14_three_bg_systems_error_signals.md - id: ARC-019 title: "REE requires staged developmental training with explicit curriculum gates." claim_type: architectural_commitment subject: developmental.curriculum_stages polarity: asserts status: provisional depends_on: - ARC-005 - ARC-006 - ARC-007 - ARC-013 - INV-010 location: docs/architecture/developmental_curriculum.md#arc-019 source: - docs/architecture/developmental_curriculum.md - id: ARC-020 title: "Offline consolidation is protected by typed authority/write boundaries." claim_type: architectural_commitment subject: sleep.protected_offline_consolidation_boundary polarity: asserts status: candidate depends_on: - ARC-011 - INV-010 - INV-021 - INV-023 location: docs/architecture/sleep.md#arc-020 source: - docs/architecture/sleep.md - docs/architecture/agency_responsibility_flow.md - docs/invariants.md - id: MECH-025 title: "Action mode prioritizes short-horizon, high-precision control via context-dependent precision modulation." claim_type: mechanism_hypothesis subject: cognitive_modes.action_doing_precision polarity: asserts status: candidate evidence_quality_note: | V3 gate cleared (2026-03-28): SD-004 (action objects), SD-005 (z_self/z_world split), and ARC-016 (dynamic precision) now implemented in ree-v3. V3 action-doing mode experiments can now proceed. V2 experiment EXQ-026 (action_doing_mode_probe, EVB-0025) FAIL -- 1 run (2026-03-08). Proximate failure: action_precision_lift = 0.0 across all seeds/conditions — E3 precision hardcoded, no dynamic channel (same root cause as Q-007). Fundamental blocker: action-doing mode requires self-attribution (SD-003) — the agent must detect it is the causal source, not read an external transition_type label. SD-003 requires E2 action-object latent (SD-004). Slope criterion: HIGH_CAUSAL (0.0211) vs LOW_CAUSAL (0.0174) — right direction, 0.0037 gap, well short of 0.05 margin; likely policy gradient density effect, not action-doing mode signal. No valid evidence in current corpus. DECOMPOSITION (2026-04-02): Original MECH-025 combined precision modulation (mechanistic, supported by Friston 2013 active inference) with responsibility linkage (architectural, philosophical bridge). Literature supports precision modulation of motor control but the responsibility linkage -- where high precision implies ethical accountability -- is a philosophical claim, not a neuroscience finding. MECH-025 now covers the mechanistic precision modulation only. Responsibility linkage decomposed to MECH-025b. EXQ-199 DIAGNOSTIC (2026-04-02, governance 2026-04-03): committed_step_count=0 both seeds. BreathOscillator creates uncommitted windows but agent never enters committed state during evaluation -- cannot measure doing-mode causal signal without committed steps. Classified as substrate limitation (not evidence against MECH-025). evidence_direction=mixed in manifest but excluded from evidence scoring for this reason. depends_on: - ARC-016 - ARC-005 - ARC-015 - ARC-021 - INV-012 - ARC-044 location: docs/architecture/modes_of_cognition.md#mech-025 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: MECH-026 title: "Ready vigilance primes restraint under high sensitivity without action." claim_type: mechanism_hypothesis subject: cognitive_modes.ready_vigilance polarity: asserts status: provisional evidence_quality_note: | AROUSAL REGULATOR NEEDED (2026-04-02): Literature (Langner & Eickhoff 2013 meta-analysis, Oken et al. 2006) establishes that vigilance = high arousal + suppressed motor output. The LC-NE inverted-U curve means optimal vigilance requires a SPECIFIC sensitivity level, not maximum. Setting precision too high produces hypervigilance (false alarms). MECH-026's ready mode needs an arousal regulator that finds optimal sensitivity, not a binary switch. MECH-093 (z_beta modulates E3 heartbeat frequency) may be the implementation mechanism: heartbeat rate IS the arousal parameter, expressed as temporal sampling frequency rather than gain scalar. See MECH-161 for the specific arousal regulator claim. MISSING CONSTRAINT -- FATIGUE (2026-04-02): Biological vigilance degrades over time as right PFC activation wanes (the vigilance decrement). REE's computational precision parameter has no fatigue and no metabolic cost. This is a missing constraint. Whether the cost of vigilance is informationally relevant (the decrement is a feature of finite-resource systems) or merely a biological limitation (REE can simply maintain steady precision) is an open question. Biology suggests the decrement is associated with attentional lapses that serve an information-gathering function (exploring vs exploiting attention allocation). ARC-044 may subsume this. depends_on: - ARC-016 - ARC-005 - MECH-093 - MECH-161 - ARC-044 location: docs/architecture/modes_of_cognition.md#mech-026 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: MECH-027 title: "Pathological modes reflect mis-tuned control-plane regimes." claim_type: mechanism_hypothesis subject: cognitive_modes.pathological_regimes polarity: asserts status: provisional depends_on: - ARC-016 - ARC-005 - ARC-015 location: docs/architecture/modes_of_cognition.md#mech-027 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: MECH-028 title: "Ethical behavior depends on mode transitions and learning preservation." claim_type: mechanism_hypothesis subject: cognitive_modes.ethical_transitions polarity: asserts status: provisional depends_on: - ARC-016 - ARC-005 - ARC-015 - INV-012 location: docs/architecture/modes_of_cognition.md#mech-028 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: MECH-029 title: "Reflective/DMN mode supports moral evaluation of hippocampally-generated replay trajectories." claim_type: mechanism_hypothesis subject: default_mode.reflective_ethics polarity: asserts status: provisional evidence_quality_note: | EXQ-065 FAIL (2026-03-23): passive observation -- agent stayed locked in committed mode throughout eval (uncommitted_step_count=0). No uncommitted state observed. Threshold (commit_threshold=0.40) never crossed by running_variance. EXQ-080 FAIL (2026-03-23): BreathOscillator active -- C1 PASS (z_world_var_ratio=1.173, sweep > inter), committed_rate=0.0 in both windows, harm_pred_std~=0 (degenerate). The oscillator creates variance differences in z_world (good) but committed_rate=0 (threshold calibration issue) and E3 harm_eval output is constant (cannot express window differentiation). REVISIT AFTER z_goal SUBSTRATE IS ACTIVE. Root cause: without a positive goal, the agent has nothing to commit TO. Natural commitment episodes require goal-directed trajectories. EXQ-074b/076 test the wanting/goal substrate; MECH-029 redesign should follow once those confirm z_goal is working. The BreathOscillator mechanism (variance difference C1 PASS) has real signal; it is the commitment detection and E3 harm_eval sensitivity that need work first. EXQ-201 FAIL (2026-04-02): improved BreathOscillator (period=50, amp=0.3, 500 warmup eps) -- uncommitted_steps=0 for both seeds [42, 123] throughout all 50 eval eps x 200 steps. Same substrate limitation as EXQ-065/080: agent cannot enter uncommitted mode under any BreathOscillator configuration tested. evidence_direction corrected to unknown (scoring_excluded: substrate_limitation). This is not evidence against MECH-029 -- the claim requires a substrate capable of generating uncommitted-mode windows, which the current V3 architecture cannot produce. All three experiments are substrate-limited and non-informative about the claim. CLARIFICATION (2026-04-02): Literature (Liu et al. 2024, Yoder & Decety 2019) suggests the DMN's moral engagement is primarily evaluative (judging outcomes of imagined scenarios), not generative (proposing the scenarios). The hippocampus generates replay trajectories; DMN/E3 evaluates them for ethical content. This actually fits REE's architecture better than the original "gated replay" framing: the hippocampal module generates candidate trajectories, E3 scores them for harm/ benefit. Title updated to reflect this generation/evaluation distinction. depends_on: - ARC-014 - ARC-005 - ARC-007 location: docs/architecture/default_mode.md#mech-029 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - id: MECH-030 title: "Sleep modes consolidate learning and ethical residue across regimes." claim_type: mechanism_hypothesis subject: sleep.modes_consolidation polarity: asserts status: provisional evidence_quality_note: | EXQ-127 FAIL (2026-03-29): superseded -- sleep not implemented in V3; post-training phase simulation does not test this claim. Manifest marked superseded. Claim remains provisional pending V4 sleep substrate implementation. depends_on: - ARC-011 - ARC-005 - ARC-007 - INV-010 - MECH-092 - MECH-094 - MECH-120 - MECH-121 - MECH-122 - MECH-123 location: docs/architecture/sleep.md#mech-030 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md - docs/thoughts/2026-02-23_some_subjective_experience_mapping.md notes: > Offline consolidation mode is multi-phase. Four sub-phases are now architecturally specified (all V4 scope): MECH-120 (SWS denoising/attractor flattening, Tononi SHY), MECH-121 (NREM SWR replay and episodic->semantic transfer, Diekelmann), MECH-122 (spindles as packaging/gating mechanism), MECH-123 (REM precision prior recalibration, Hobson/Friston). All sub-phases require MECH-094 hypothesis tag as hard prerequisite to prevent harm trace over-amplification (MECH-124). V3 prerequisites for V4 consolidation: MECH-092 quiescent replay validated; MECH-094 hypothesis tag working; D_eff monitoring (MECH-113) with stable setpoint; z_goal representation (MECH-112, 116) validated; multi-rate execution (SD-006, EXQ-052b PASS); ThetaBuffer bidirectional design (currently waking-direction only, MECH-122 requires sleep-direction). See docs/architecture/v3_v4_transition_boundary.md. Literature: Tononi & Cirelli 2006 (SHY), Diekelmann & Born 2010 (SWR consolidation), Hobson & Friston 2012 (REM precision recalibration), Walker & Stickgold 2004 (maladaptive consolidation / PTSD model -> MECH-124). Evidence confidence: 0.79-0.88. Updated 2026-03-23 to add sub-phase claims. - id: MECH-031 title: "Derived social tags and empathy coupling via control-plane knobs." claim_type: mechanism_hypothesis subject: social.other_selflike_empathy_coupling polarity: asserts status: provisional depends_on: - ARC-010 - ARC-005 - ARC-017 - ARC-015 - ARC-006 - INV-005 location: docs/architecture/social.md#mech-031 source: - docs/thoughts/2026-02-09_empathy.md - id: MECH-032 title: "OTHER_SELFLIKE detection is biased toward high recall to avoid empathy false negatives." claim_type: mechanism_hypothesis subject: social.other_selflike_high_recall_bias polarity: asserts status: provisional depends_on: - MECH-031 - ARC-010 location: docs/architecture/social.md#mech-032 source: - docs/architecture/social.md - id: MECH-033 title: "E2 forward-prediction kernels seed hippocampal rollouts." claim_type: mechanism_hypothesis subject: hippocampus.kernel_chaining_interface polarity: asserts status: active lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree v3_pending: false evidence_quality_note: | V2 experiment EXQ-022 (kernel_chaining_interface, EVB-0016) FAIL — run twice (2026-03-08, 2026-03-15), consistent result. WITH_CHAIN (full E2→hippocampus→E3 pipeline) only 1.8% better than NO_CHAIN (random action selection); random agent improves faster on slope. Root cause: V2 E2 produces z_gamma sensory transitions, not action-consequence objects. Chaining sensory predictions into rollouts provides negligible signal to E3 — the interface exists but passes the wrong type across it. Structural failure, not a tuning issue. V3 EXQ-055 PASS (2026-03-20): action-object chaining vs self-only vs random — AO reduces harm 67× vs self-only (harm_ao=0.00147 vs harm_self=0.0684), cal_gap_ao=0.734 vs cal_gap_self=0.021. All 5 criteria met. V3 z_world action-object E2 is the correct primitive. v3_pending cleared 2026-03-20. EXQ-124 FAIL (2026-03-28): harm_rate_CHAIN=0.621 vs ABLATED=0.057 (~10x worse with kernel chain). Training calibration issue: 400-episode warmup insufficient, E2 world_forward not well trained, terrain prior not calibrated for kernel-chain conditions. Not fundamental failure of the claim -- prior EXQ-055 PASS remains valid (different baseline). Experiment design problem. Literature pull on hippocampal planning queued (Pfeiffer & Foster 2013; Wikenheiser & Redish 2015). EXQ-171 PASS 4/4 x2 (2026-03-30, two independent runs): kernel chain confirmed. harm_rate_CHAIN=0.00021 vs ABLATED=0.0613 (290x safer). world_forward_r2=0.984. harm_reduction_delta=0.0611 >> 0.01 threshold. Both seeds, both runs. EXQ-124 training-budget FAIL superseded -- 600-ep warmup sufficient. EXQ-055 + EXQ-171x2 = three clean PASS entries. Evidence_direction: supports. EXQ-184 PASS 3/3 seeds (2026-04-04): Independent replication. harm_rate_CHAIN=0.00021 vs NO_CHAIN=0.0608 (290x safer). world_forward_r2=0.987. harm_reduction_delta=0.0606. All 4 criteria met across all seeds. Third clean PASS; evidence strong. depends_on: - ARC-018 - ARC-002 - ARC-001 - ARC-005 - ARC-021 location: docs/architecture/hippocampal_systems.md#mech-033 source: - docs/thoughts/2026-02-09_e2_hpc_interface.md - id: MECH-034 title: "Viability mapping updates are distinct from residue curvature updates." claim_type: mechanism_hypothesis subject: residue.viability_vs_curvature_updates polarity: asserts status: provisional depends_on: - ARC-013 - ARC-018 location: docs/architecture/residue_geometry.md#mech-034 source: - docs/thoughts/2026-02-09_viability_mapping_vs_residue.md - id: MECH-035 title: "VALENCE is vector-valued and ranked without scalar collapse." claim_type: mechanism_hypothesis subject: valence.vector_ranking polarity: asserts status: candidate depends_on: - ARC-017 - ARC-003 - ARC-005 location: docs/architecture/sensory_stream_tags.md#mech-035 source: - docs/thoughts/2026-02-09_valence_vector.md - id: MECH-036 title: "Other-harm triggers veto only under high-certainty catastrophic outcomes." claim_type: mechanism_hypothesis subject: social.other_harm_veto_threshold polarity: asserts status: provisional depends_on: - MECH-031 - ARC-005 - INV-005 location: docs/architecture/social.md#mech-036 source: - docs/thoughts/2026-02-09_other_harm_gating.md - id: MECH-037 title: "Papez-like loop provides provenance gating and reality filtering." claim_type: mechanism_hypothesis subject: provenance.papez_like_reality_filtering polarity: asserts status: provisional depends_on: - ARC-007 - ARC-018 - ARC-003 - ARC-005 location: docs/architecture/papez_circuit.md#mech-037 source: - docs/thoughts/2026-02-09_papez_circuit_reality_filtering.md - id: MECH-038 title: "Arcuate-like sequence-to-motor channel nudges language emergence." claim_type: mechanism_hypothesis subject: language.arcuate_fasciculus_nudges polarity: asserts status: candidate depends_on: - ARC-009 - ARC-002 - ARC-018 - ARC-005 location: docs/architecture/arcuate_fasciculus.md#mech-038 source: - docs/thoughts/2026-02-09_arcuate_fasciculus_language_nudges.md - id: MECH-039 title: "Modes are stable regions in control-channel space, not separate modules." claim_type: mechanism_hypothesis subject: control_plane.channel_mode_landscape polarity: asserts status: provisional depends_on: - ARC-005 - ARC-016 - MECH-019 location: docs/architecture/control_plane.md#mech-039 source: - docs/architecture/control_plane.md - id: MECH-040 title: "Safety baseline and volatility are distinct control channels for arousal/readiness." claim_type: mechanism_hypothesis subject: control_plane.safety_baseline_volatility polarity: asserts status: provisional lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree evidence_quality_note: | No experimental evidence in current genuine corpus. All prior entries (9 runs, synthetic scaffolding) originated from ree-v2/ree-experiments-lab (archived 2026-02-26) and have been removed from the index. The demotion recommendation in promotion_demotion_recommendations.md (conflict_ratio=0.889, conf=0.55) is stale — it was computed from those invalid entries, which no longer exist. Retain provisional (adjudication retain_ree is valid as architectural decision) pending genuine ree-v1-minimal experiments. Do not act on demotion recommendation until genuine experimental evidence is available for this claim. depends_on: - ARC-005 - MECH-005 - MECH-019 location: docs/architecture/control_plane.md#mech-040 source: - docs/architecture/control_plane.md - id: MECH-041 title: "Affective expression broadcasts control-plane regime to reduce social prediction load." claim_type: mechanism_hypothesis subject: social.affective_expression_mode_broadcast polarity: asserts status: provisional depends_on: - ARC-010 - ARC-005 - ARC-016 - MECH-039 location: docs/architecture/social.md#mech-041 source: - docs/thoughts/2026-02-10_affective_expression_mode_broadcast.md - id: MECH-042 title: "Telemetry exposure channels report internal control-plane state for diagnostics." claim_type: mechanism_hypothesis subject: control_plane.telemetry_exposure_channels polarity: asserts status: candidate depends_on: - ARC-005 - MECH-039 - MECH-040 location: docs/architecture/control_plane.md#mech-042 source: - docs/thoughts/2026-02-10_control_plane_telemetry.md - id: MECH-043 title: "Dopamine-like modulation of precision-weighting for unsigned prediction errors." claim_type: mechanism_hypothesis subject: precision.dopamine_weighting polarity: asserts status: provisional depends_on: - ARC-005 - INV-008 - MECH-003 location: docs/architecture/precision_control.md#mech-043 source: - docs/thoughts/2026-02-10_dopamine_precision_weighting.md - id: MECH-044 title: "Hippocampal systems participate in relational binding and comparison." claim_type: mechanism_hypothesis subject: entities.hippocampal_relational_binding polarity: asserts status: provisional depends_on: - ARC-006 - ARC-007 - ARC-004 location: docs/architecture/entities_and_binding.md#mech-044 source: - docs/thoughts/2026-02-10_hippocampal_relational_binding.md - id: MECH-045 title: "Object-file-like buffers provide minimal entity persistence across time." claim_type: mechanism_hypothesis subject: entities.object_file_persistence polarity: asserts status: provisional depends_on: - ARC-006 - ARC-004 - INV-002 location: docs/architecture/entities_and_binding.md#mech-045 source: - docs/thoughts/2026-02-10_object_file_persistence.md - id: MECH-046 title: "Amygdala analogue updates control-plane mode priors from fast salience classification." claim_type: mechanism_hypothesis subject: control_plane.amygdala_mode_priors polarity: asserts status: provisional lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree evidence_quality_note: | No experimental evidence in current genuine corpus. All prior entries (9 runs, synthetic scaffolding) originated from ree-v2/ree-experiments-lab (archived 2026-02-26) and have been removed from the index. The demotion recommendation in promotion_demotion_recommendations.md (conflict_ratio=0.889, conf=0.55) is stale — it was computed from those invalid entries, which no longer exist. Retain provisional (adjudication retain_ree is valid as architectural decision) pending genuine ree-v1-minimal experiments. Do not act on demotion recommendation until genuine experimental evidence is available for this claim. depends_on: - ARC-005 - MECH-039 location: docs/architecture/control_plane.md#mech-046 source: - docs/thoughts/2026-02-11_amygdala.md - id: MECH-047 title: "Pre-commitment mode manager commits with hysteresis and switching costs." claim_type: mechanism_hypothesis subject: control_plane.precommitment_mode_manager polarity: asserts status: provisional depends_on: - ARC-005 - MECH-046 location: docs/architecture/mode_manager.md#mech-047 source: - docs/thoughts/2026-02-11_amygdala.md - docs/thoughts/2026-02-11_some_control_plane_maths_hypotheses.md - id: MECH-048 title: "mu/kappa stability overlays modulate mode entropy and switching pressure." claim_type: mechanism_hypothesis subject: control_plane.mu_kappa_stability_overlays polarity: asserts status: provisional depends_on: - ARC-005 - MECH-039 location: docs/architecture/control_plane.md#mech-048 source: - docs/thoughts/2026-02-11_opioid_receptors.md - docs/thoughts/2026-02-11_some_control_plane_maths_hypotheses.md - id: MECH-049 title: "Temporal phase compartmentalisation preserves ethical constraint independence." claim_type: mechanism_hypothesis subject: temporal.phase_compartmentalisation polarity: asserts status: candidate depends_on: - ARC-008 - ARC-003 - ARC-005 location: docs/architecture/temporal_dynamics.md#mech-049 source: - docs/thoughts/2026-02-11_phase_separation.md - id: MECH-050 title: "Functional locality supports attribution without requiring anatomical columns." claim_type: mechanism_hypothesis subject: entities.functional_locality_without_columns polarity: asserts status: candidate depends_on: - ARC-006 - ARC-004 location: docs/architecture/entities_and_binding.md#mech-050 source: - docs/thoughts/2026-02-11_columns.md - id: MECH-051 title: "Oxytocin/vasopressin analogues modulate relational topology and mode priors." claim_type: mechanism_hypothesis subject: social.relational_topology_modulation polarity: asserts status: candidate depends_on: - ARC-010 - ARC-005 - ARC-006 location: docs/architecture/social.md#mech-051 source: - docs/thoughts/2026-02-11_oxytocin_vasopressin.md - id: MECH-052 title: "Prolactin analogue stabilises care-investment persistence." claim_type: mechanism_hypothesis subject: social.care_investment_persistence polarity: asserts status: candidate depends_on: - ARC-010 - MECH-051 - MECH-048 location: docs/architecture/social.md#mech-052 source: - docs/thoughts/2026-02-11_prolactin.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: MECH-053 title: "Habenula-like aversive PE gate suppresses commitment under negative spikes." claim_type: mechanism_hypothesis subject: control_plane.habenula_aversive_gate polarity: asserts status: provisional depends_on: - ARC-005 - MECH-039 - MECH-043 location: docs/architecture/control_plane.md#mech-053 source: - docs/thoughts/2026-02-11_habenula_signed_pe.md - id: MECH-054 title: "Signed harm/benefit prediction-error precision channels remain distinct." claim_type: mechanism_hypothesis subject: precision.signed_harm_benefit polarity: asserts status: provisional depends_on: - ARC-005 - ARC-017 - MECH-043 - INV-008 location: docs/architecture/control_plane.md#mech-054 source: - docs/thoughts/2026-02-11_habenula_signed_pe.md - id: MECH-055 title: "Affective channel separation keeps hedonic tone, valence, and signed PE distinct." claim_type: mechanism_hypothesis subject: affect.channel_separation polarity: asserts status: candidate depends_on: - ARC-005 - MECH-048 - MECH-054 - MECH-035 location: docs/architecture/control_plane.md#mech-055 source: - docs/architecture/control_plane.md # JEPA flags removed 2026-02-26 — residue trajectory placement is REE-native. - id: MECH-056 title: "Residue should constrain trajectories before distorting core representations." claim_type: mechanism_hypothesis subject: residue.trajectory_first_placement polarity: asserts status: provisional lifecycle_stage: adjudicated adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: hybridize adjudication_mode: external_precedence_allowed allowed_conflict_outcomes: - retain_ree - hybridize - adopt_jepa_structure - retire_ree_claim cascade_on_outcomes: - adopt_jepa_structure - retire_ree_claim dependency_reopen_status: candidate evidence_quality_note: | Prior 42 synthetic runs (ree-v2/ree-experiments-lab, archived 2026-02-26) are invalid. Genuine ree-v1-minimal experiment completed 2026-02-26 (EVB-0039, residue_trajectory_placement): PASS. TRAJECTORY-WIDE last-Q harm 0.724 vs ENDPOINT-ONLY 0.813 (11% improvement, 3 seeds). Intermediate residue mass confirmed non-zero (mean 3.48 across seeds). Both criteria met: path-spread (residue localised along planned path) and harm-avoidance (trajectory-wide accumulation reduces harm vs endpoint-only). Supports φ(z) terrain concept: harm sites are localised along planned trajectories, not only post-hoc at terminal steps. Promoted candidate → provisional 2026-02-26. Re-test on more complex substrate before active. depends_on: - ARC-013 - ARC-018 - ARC-003 - ARC-004 - MECH-034 - MECH-060 - MECH-062 location: docs/architecture/residue_geometry.md#mech-056 source: - docs/architecture/residue_geometry.md - docs/thoughts/2026-02-12_TRAJECTORY-RESIDUE-VS-REPRESENTATIONAL-DISTORTION.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md # MECH-057 split 2026-03-15: action-loop and thought-loop gates are distinct mechanisms # requiring different substrates and different V-phase targets. MECH-057a retains the V1/V2 # evidence and is V2-testable with multi-step substrate. MECH-057b is explicitly V3-scoped. - id: MECH-057a title: "Committed action sequences suppress sensory precision reweighting until execution completes." claim_type: mechanism_hypothesis subject: agentic_extension.action_loop_completion_gate polarity: asserts status: candidate lifecycle_stage: adjudicated adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-25T14:59:46.253616Z adjudication_decision_status: applied adjudication_recommendation: hybridize evidence_quality_note: | Prior 10 synthetic runs (ree-v2/ree-experiments-lab, archived 2026-02-26) are invalid. Genuine ree-v1-minimal experiment completed 2026-02-26 (EVB-0042, control_completion_requirement): FAIL (informative baseline — effect direction correct, magnitude sub-threshold). FULL last-Q harm 0.879 | NO_ATTRIBUTION 0.896 (+1.9%) | NO_GATING 0.919 (+4.5%). Degradation threshold 1.1 (10%): neither ablation met it (1.019x and 1.046x respectively). Effect ordering is architecturally correct: FULL < NO_ATTRIBUTION < NO_GATING — completion gating matters more than attribution at this scale (4.5% vs 1.9%), consistent with theory. V2 re-test (Attribution Completion Gating — V2 Redesign, 2026-03-15): FAIL. V2 substrate still uses single-step grid actions — atomic action structure persists in base CausalGridWorld. The action-loop gate has no "action in progress" state to protect. INTERPRETATION: all prior FAILs are substrate-limited, not architecture falsifications. SCOPE REVISION RESOLVED (2026-03-15): MECH-057 split into MECH-057a (this claim, action loop) and MECH-057b (thought-loop trajectory promotion gate, V3-scoped). The two gates are mechanistically distinct and require different substrates. SUBSTRATE FIX IMPLEMENTED (2026-03-15): CausalGridWorld extended with subgoal_mode=True (3-waypoint committed sequences W0→W1→W2; entity type 6; timeout=20 steps). REEAgent updated: action_loop_gate_enabled flag + sequence_in_progress routing in generate_trajectories(); update_residue(owned=bool) for attribution ablation. EXQ-020 designed and smoke-tested 2026-03-15 (directionally PASS at 5 eps / 1 seed). Full run (5 seeds × 250 episodes × 3 conditions) queued in ree-v2/experiment_queue.json. Will update to provisional on EXQ-020 PASS. CLAIM REFRAMING (2026-03-15): Heartbeat architecture analysis (ARC-023, MECH-090) reveals the experiment modelled E3 as updating only at completion (binary gate), not as beta-gated propagation within a continuous update stream. The correct mechanism is: completion events are the principal policy-update opportunities (beta drops, E3 model propagates to action selection) within an otherwise continuous E3 internal update stream. The V2 FAIL pattern (ATTRIBUTION_BLIND comparable to COMPLETION_GATED) is consistent with this reframing: continuous updating without any beta gate is closer to correct architecture than episodic gating. Subject updated to reflect: completion events as policy-propagation gates, not as exclusive E3 update triggers. Full test requires ARC-023 + MECH-090 (V3 substrate, SD-006). EXQ-139 FAIL (2026-03-29): mean_committed_seq_len_on=0.0 -- committed sequences never generated; gate never triggered at this training scale; harm_rate_GATE_ON=harm_rate_ABLATED= 0.7244. Substrate issue, not scientific failure. Hold pending investigation of why committed sequences fail to generate at current training budget. FRAMING CLARIFICATION (2026-04-02): Literature (Schmidt et al. 2019, Leventhal et al. 2012) distinguishes motor output suppression (beta blocks movement initiation) from sensory precision reweighting suppression (beta blocks precision updates during committed sequences). MECH-057a claims the latter -- beta suppresses SENSORY REWEIGHTING, not motor output. MECH-090 (beta gates E3->action propagation, not E3 internal updating) is probably the more correct framing and is better supported by electrophysiology. However, the alternative framing (beta suppresses precision updates themselves, not just their propagation) is worth testing as lower priority once substrate supports it. Title updated from "precision updates" to "sensory precision reweighting." depends_on: - ARC-015 - ARC-005 - ARC-003 - ARC-004 - INV-012 - ARC-023 - MECH-090 location: docs/architecture/agency_responsibility_flow.md#mech-057a source: - docs/thoughts/2026-02-13_LeCun_developed_lots_of_REE.md - docs/thoughts/2026-02-13_subjective_experience_pre_post_commit.md - docs/thoughts/2026-02-23_some_subjective_experience_mapping.md # Reframed 2026-02-26 — removed JEPA framing. Subject renamed to REE-native formulation. # Prior evidence (156 entries) was generated by synthetic repos (ree-v2, ree-experiments-lab, # archived 2026-02-26) and is invalid. Claim reset for genuine ree-v1-minimal experiments. # Split from MECH-057 2026-03-15 — scope revision resolved. - id: MECH-057b title: "Hippocampal sequence completion must be verified before candidates are eligible for E3 selection." claim_type: mechanism_hypothesis subject: agentic_extension.thought_loop_trajectory_promotion_gate polarity: asserts status: candidate lifecycle_stage: candidate implementation_phase: v3 literature_evidence: | Pfeiffer & Foster (2013) PMID 23594744 [DOI: 10.1038/nature12112]: Hippocampal place-cell sequences depict future paths to remembered goals before goal-directed navigation. Sequences are generated, checked for goal-reach plausibility, then emitted — this is the planning/candidacy gate in action. The mechanism confirms that trajectory candidacy is decided in hippocampus before proposals reach downstream planning stages. Central reference for MECH-057b. Diba & Buzsáki (2007) PMID 17828259 [DOI: 10.1038/nn1961]: Forward and reverse hippocampal place-cell sequences during sharp-wave ripples. Forward sequences are anticipatory (planning); reverse sequences are evaluative (completion verification). The switch from reverse to forward replay at run start is a completion marker — the previous sequence ran to its end and was evaluated. Supports the idea that sequence completion is a discrete detectable event. Lansink et al. (2009) PMID 19688032 [DOI: 10.1371/journal.pbio.1000173]: Hippocampus leads ventral striatum in joint replay. Place cells fire before reward-correlated striatal cells. Directionality confirmed: hippocampal completion signals propagate forward to BG, not the reverse. Establishes the anatomical direction of the candidacy→propagation gate coupling (see MECH-105, ARC-028). evidence_quality_note: | No genuine experiments. This claim is explicitly V3-scoped. The thought-loop trajectory promotion gate requires HippocampalModule to implement a feedback path that suppresses trajectory candidates from being promoted to E3 consideration before hippocampal sequence completion is verified. This path is not present in V1 or V2. V3 primary scope: full HippocampalModule trajectory promotion policy and E3 complex commitment gate (per V2 spec Step 2.5 exit criteria and V3 control completion focus). Split from MECH-057 2026-03-15 -- see MECH-057a for action-loop evidence and V1/V2 history. Tagging correction (2026-03-22): EXQ-048/048b/059/060 were previously tagged as MECH-057b evidence, but all of these experiments tested the BG beta propagation gate (MECH-090), not the hippocampal candidacy gate. The tagging error arose because EXQ-048 was designed for MECH-057b but the implementation bypassed BetaGate entirely (broken instrumentation); subsequent iterations fixed the routing and shifted to testing MECH-090, but carried the MECH-057b tag forward without re-evaluation. All affected result JSONs and manifests have been corrected. MECH-057b has zero genuine experimental evidence. The governance confidence score (0.66) prior to this fix was an artefact of the tagging error. First genuine experiment needed: a dedicated V3 experiment instrumenting HippocampalModule's trajectory emission directly -- measuring whether partial sequences are suppressed vs completed sequences are promoted to E3 evaluation. This requires HippocampalModule trajectory promotion policy to be implemented first (V3 scope). See ARC-028 for the coupling between the hippocampal completion signal and BetaGate. depends_on: - ARC-015 - ARC-005 - ARC-003 - ARC-004 - INV-012 - Q-019 - ARC-020 location: docs/architecture/agency_responsibility_flow.md#mech-057b source: - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - docs/thoughts/2026-03-14_three_bg_systems_error_signals.md - id: MECH-058 title: "Slow target-anchor dynamics stabilize E1/E2 representations via functional rate separation." claim_type: mechanism_hypothesis subject: latent_stack.e1_e2_timescale_separation polarity: asserts status: retired lifecycle_stage: superseded superseded_by: MECH-069 superseded_at_utc: 2026-03-15T00:00:00Z supersession_note: | MECH-058 framed the E1/E2 distinction as a *timescale* difference (learning rate separation). This framing was a stepping stone but is now understood to be incorrect. The deeper and more precise claim is MECH-069: E1, E2, and E3 train on *incommensurable error signals* (sensory prediction error / motor-sensory error / harm+goal error). These are not merely different timescales — they are functionally distinct domains that cannot be merged without misattributing credit. EXQ-019 (V2 redesign, 2026-03-15) FAILED: learning-rate separation produced no benefit, as expected — the test was asking the wrong question. Testing LR ablation cannot reveal functional-domain incommensurability. MECH-069 supersedes this claim and can only be properly tested in V3 with the z_self/z_world latent split (SD-005). adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-25T16:00:13.142266Z adjudication_decision_status: applied evidence_quality_note: | V1 EXQ-002 (e1_e2_timescale_ablation): FAIL. V2 EXQ-019 (terrain_timescale, 2026-03-15): FAIL. Both FAILs consistent with supersession: the test was ill-posed. LR separation ≠ functional domain separation. See MECH-069 for the correct claim. depends_on: - ARC-001 - ARC-002 - ARC-004 - ARC-015 - MECH-057 location: docs/architecture/agency_responsibility_flow.md#mech-058 source: - docs/architecture/agency_responsibility_flow.md # Promoted 2026-02-26 — retain_ree adjudication was applied 2026-02-15 - id: MECH-059 title: "Confidence channel (uncertainty-derived precision) must remain distinct from residual error." claim_type: mechanism_hypothesis subject: precision.confidence_channel_separate_from_prediction_error polarity: asserts status: active lifecycle_stage: active adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree adjudication_mode: external_precedence_allowed allowed_conflict_outcomes: - retain_ree - hybridize - adopt_jepa_structure - retire_ree_claim cascade_on_outcomes: - adopt_jepa_structure - retire_ree_claim dependency_reopen_status: candidate evidence_quality_note: "Prior 38 synthetic runs invalidated (ree-v2/ree-experiments-lab, archived 2026-02-26). Genuine ree-v1-minimal experiment completed 2026-02-26 (control_plane_precision_separation, EVB-0037): PASS — |corr(score_dispersion, PE)| = 0.067 (<0.3), SEPARATED harm 0.663 vs MERGED 0.875. Both mechanistic criteria met." depends_on: - ARC-005 - ARC-004 - ARC-015 - MECH-054 - MECH-057 location: docs/architecture/agency_responsibility_flow.md#mech-059 source: - docs/architecture/agency_responsibility_flow.md # JEPA flags removed 2026-02-26 — dual error channel concept is REE-native. # pending_design: requires pure REE implementation without JEPA substrate proxy. - id: MECH-060 title: "Pre-commit simulation and post-commit realized-error channels must stay responsibility-distinct via write-boundary enforcement." claim_type: mechanism_hypothesis subject: commitment.dual_error_channels_pre_post_commit polarity: asserts status: provisional lifecycle_stage: adjudicated evidence_quality_note: | Prior 137 synthetic runs (ree-v2/ree-experiments-lab, archived 2026-02-26) are invalid. pending_design cleared 2026-02-27: both channels exist natively in ree-v1-minimal — sim_error = E2 harm_predictions on selected trajectory (pre-commit, before env.step), realized_error = env harm returned by env.step (post-commit). Write-locus separation is enforced by construction (E1 trains only on actual observations via compute_prediction_loss, residue accumulates only via update_residue with post-commit harm). MECH-061 PASS (EXQ-003) confirms channels are genuinely distinct (|corr|=0.114) and separation is beneficial. Genuine ree-v1-minimal experiment completed 2026-02-27 (EVB-0043, write_locus_contamination, EXQ-005): PASS. 200 eps × 3 seeds. FULL last-Q harm 0.929 | CONT_DURABLE 0.915 | CONT_RESIDUE 0.593. Criterion 1 (residue inflation): CONT_RESIDUE total_residue 8520 > FULL 186 × 1.1 = 205 — MET. Criterion 2 (harm ordering): FULL 0.929 ≤ max_cont 0.915 × 1.05 = 0.960 — MET. Write-locus contamination is detectable: pre-commit residue pollution inflates the residue field 46× vs clean. FULL condition harm ordering is sustained, confirming write-locus separation is architecturally load-bearing. Promoted candidate → provisional 2026-02-27. adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-25T16:33:23.830904Z adjudication_decision_status: applied adjudication_recommendation: hybridize adjudication_mode: external_precedence_allowed allowed_conflict_outcomes: - retain_ree - hybridize - adopt_jepa_structure - retire_ree_claim cascade_on_outcomes: - adopt_jepa_structure - retire_ree_claim dependency_reopen_status: candidate depends_on: - ARC-003 - ARC-005 - ARC-015 - INV-012 - MECH-057 location: docs/architecture/agency_responsibility_flow.md#mech-060 source: - docs/architecture/agency_responsibility_flow.md - docs/thoughts/2026-02-13_subjective_experience_pre_post_commit.md - docs/thoughts/2026-02-23_some_subjective_experience_mapping.md - docs/thoughts/2026-02-24_determinism_action_gating_boundary.md # Promoted 2026-02-26 — retain_ree adjudication was applied 2026-02-15 - id: MECH-061 title: "Commit-boundary token reclassifies pre-commit vs post-commit error routing." claim_type: mechanism_hypothesis subject: commitment.boundary_token_error_reclassification polarity: asserts status: active lifecycle_stage: active adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied evidence_quality_note: | Prior 8 synthetic runs (ree-v2/ree-experiments-lab, archived 2026-02-26) are invalid. Genuine ree-v1-minimal experiment completed 2026-02-26 (EVB-0041, commitment_boundary_validation): PASS. WITH-BOUNDARY last-Q harm 0.805 vs BLENDED 0.935 (3 seeds). Mean |pre_post_corr| = 0.044 (well below 0.7 threshold). Both criteria met: distinct-signals (|corr| < 0.7 confirms pre-commit E2 predictions and post-commit realized harm carry distinct information) and boundary-helps (clean separation enables better policy learning than blending). The commit boundary is operationally load-bearing at ree-v1-minimal scale — not a cosmetic architectural feature. Validates retain_ree adjudication. Status remains active. adjudication_recommendation: retain_ree depends_on: - ARC-003 - ARC-015 - INV-012 - MECH-060 location: docs/architecture/e3.md#mech-061 source: - docs/architecture/e3.md - docs/thoughts/2026-02-15_basal_ganglia.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - docs/thoughts/2026-02-24_determinism_action_gating_boundary.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: MECH-062 title: "E3 uses coordinated tri-loop gating (motor, cognitive-set, motivational) as the pre-commit eligibility layer." claim_type: mechanism_hypothesis subject: commitment.tri_loop_gate_coordination polarity: asserts status: stable depends_on: - ARC-003 - ARC-005 - MECH-061 location: docs/architecture/e3.md#mech-062 source: - docs/architecture/e3.md - docs/thoughts/2026-02-15_basal_ganglia.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - docs/thoughts/2026-02-13_subjective_experience_pre_post_commit.md - docs/thoughts/2026-02-23_some_subjective_experience_mapping.md - docs/thoughts/2026-02-24_prefrontal_primitives.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: MECH-063 title: "Control plane retains orthogonal tonic/phasic axes rather than collapsing into one scalar." claim_type: mechanism_hypothesis subject: control_plane.orthogonal_axes_tonic_phasic polarity: asserts status: provisional depends_on: - ARC-005 - MECH-039 - MECH-040 - MECH-055 location: docs/architecture/control_plane.md#mech-063 source: - docs/architecture/control_plane.md - docs/thoughts/2026-02-15_basal_ganglia.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - docs/thoughts/2026-02-13_subjective_experience_pre_post_commit.md - docs/thoughts/2026-02-23_some_subjective_experience_mapping.md - id: MECH-064 title: "Typed authority/control-store separation blocks direct exteroceptive writes into policy and identity stores." claim_type: mechanism_hypothesis subject: control_plane.typed_authority_control_store_separation polarity: asserts status: candidate depends_on: - ARC-005 - ARC-003 - ARC-015 - INV-014 - INV-007 - MECH-062 location: docs/architecture/control_plane_signal_map.md#mech-064 source: - docs/architecture/control_plane_signal_map.md - docs/thoughts/17-02-26_necessary_separations_based_on_considering-prompt_injection.md - id: MECH-065 title: "Reality-coherence conflict lane modulates loop precision and commitment thresholds before execution lock-in." claim_type: mechanism_hypothesis subject: control_plane.reality_coherence_conflict_lane polarity: asserts status: candidate depends_on: - ARC-005 - ARC-007 - ARC-018 - MECH-037 - MECH-054 - MECH-062 notes: > Literature grounding (2026-03-29, Q-018 lit pull): Sterzer et al. (Front Hum Neurosci 2016) reframes the RC-conflict threshold as a precision RATIO rather than an absolute signal strength -- thought insertion and misattribution arise when the ratio of internal-model precision to external-authority precision is miscalibrated. Howes & Murray (Biol Psych 2020) show that early adverse experience systematically lowers this ratio threshold (aberrant salience), and that the CHR-to-psychosis transition is the hysteresis zone corresponding to a threshold that is both too low to resist spoofing and too unstable to self-correct. Damiani et al. (meta-analysis 2022) confirms internal source monitoring is the most vulnerable calibration target (SMD ~0.5-1.0 deficit). Implementation implication: the RC-conflict lane threshold should be implemented as ratio(E1_precision, authority_signal_precision), not an absolute authority-signal magnitude. This directly addresses Q-018's calibration question. Clinical analogue: authority_spoofing_low_threshold -> psychosis/paranoia; over_suppression_high_threshold -> rigid_hypervigilance/anhedonia. Clinical analogues are tracked as notes rather than a separate field -- the psychiatric symptom mapping is evidence for the architectural claim, not a separate claim type. location: docs/architecture/control_plane_signal_map.md#mech-065 source: - docs/architecture/control_plane_signal_map.md - docs/thoughts/2026-02-17_control_plane_update.md - docs/thoughts/17-02-26_necessary_separations_based_on_considering-prompt_injection.md - id: MECH-066 title: "Pre-commit and post-commit channels may share representations but must stay separated at durable write boundaries." claim_type: mechanism_hypothesis subject: agency.boundary_conditioned_channel_separation polarity: asserts status: candidate depends_on: - ARC-015 - MECH-060 - MECH-061 - INV-021 location: docs/architecture/agency_responsibility_flow.md#mech-066 source: - docs/architecture/agency_responsibility_flow.md - docs/invariants.md - id: MECH-067 title: "A machine-checkable phase/store/actor permission matrix is required to enforce commit-boundary write rules." claim_type: mechanism_hypothesis subject: agency.write_locus_permission_matrix polarity: asserts status: provisional evidence_quality_note: | Co-tested with MECH-060 in genuine ree-v1-minimal experiment 2026-02-27 (EVB-0043, write_locus_contamination, EXQ-005): PASS. The write-locus permission matrix is operationally validated: pre-commit and post-commit write loci are genuinely distinct, and contaminating the residue write locus (pre-commit) inflates the residue field (8520× vs FULL 186) without corresponding harm reduction — confirming that post-commit is the correct write locus for durable harm attribution. Promoted candidate → provisional 2026-02-27. depends_on: - MECH-060 - MECH-061 - MECH-066 - INV-020 - INV-021 location: docs/architecture/agency_responsibility_flow.md#mech-067 source: - docs/architecture/agency_responsibility_flow.md - docs/architecture/control_plane_signal_map.md - id: MECH-068 title: "Consolidation selectivity lives in the consolidation operator, not in E1 feature basis." claim_type: mechanism_hypothesis subject: representation.consolidation_selectivity polarity: asserts status: candidate depends_on: - ARC-001 - ARC-003 - ARC-005 - INV-014 - MECH-062 - MECH-063 location: docs/architecture/compact_consolidation_principle.md#mech-068 source: - docs/architecture/compact_consolidation_principle.md - evidence/literature/visual_cortex_compact_models/entries/2026-02-26_cowley_2023_compact_v4/record.json # Demoted 2026-02-26 — accepted governance recommendation. Evidence 4:4 split (conflict_ratio 1.0, # confidence 0.503). Insufficient signal to hold as active question. Needs discriminative experiment. - id: Q-012 title: "Can latent predictive world models stay agentically stable without explicit REE-like control constraints?" claim_type: open_question subject: latent_predictive_models.control_completion_falsifiability polarity: asks status: candidate lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree evidence_quality_note: | All 8 prior experimental entries are synthetic (ree-v2, _toyenv_internal_minimal run_ids, archived 2026-02-26). The 4:4 supports/weakens split and 0.503 confidence score are entirely synthetic artefacts -- do NOT act on any demote recommendation based on this data. 3 literature entries are valid (2 supports, 1 mixed) and insufficient alone to resolve the question. Retain candidate status pending genuine ree-v1-minimal runs. EXQ-152 FAIL/mixed (2026-03-29): REE_FULL vs PREDICTIVE_ONLY delta_harm=0; informative null at current training scale. Significant training budget expected to be required before meaningful differences emerge between full REE and predictive-only baselines. EXQ-193 FAIL/mixed (2026-04-04): REE_FULL vs PREDICTIVE_ONLY vs NO_LEARNING (3 conditions). C1 PASS 3/3: REE_FULL outperforms PREDICTIVE_ONLY by 7.5 harm units (11.96 vs 19.47); PREDICTIVE_ONLY harm INCREASES over training while REE_FULL improves -- strong directional signal that control constraints matter. C2 FAIL: REE_FULL does not beat NO_LEARNING floor in all seeds (2/3 pass). evidence_direction=mixed. The C1 PASS is the primary Q-012 signal: pure predictive (no control constraints) destabilises over time while REE does not. This is the first experiment showing a genuine positive REE-vs-predictive gap at current training scale. Consistent with retain_ree adjudication. depends_on: - MECH-057 - ARC-015 - ARC-005 - ARC-004 location: docs/architecture/agency_responsibility_flow.md#q-012 source: - docs/thoughts/2026-02-13_LeCun_developed_lots_of_REE.md # Q-013 retired 2026-03-29 -- JEPA integration decoupled from REE. # The uncertainty calibration question is moot without a JEPA substrate. # The REE-native precision routing question is now addressed by MECH-059 directly. - id: Q-013 title: "Can deterministic JEPA plus derived dispersion match explicit stochastic uncertainty heads for REE precision routing?" claim_type: open_question subject: uncertainty.deterministic_vs_stochastic_jepa_calibration polarity: asks status: legacy lifecycle_stage: retired adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-25T16:51:50.720541Z adjudication_decision_status: applied adjudication_recommendation: hybridize evidence_quality_note: > JEPA integration dropped from REE architecture 2026-03-29. This question is moot. All prior experimental evidence (jepa_uncertainty_channels, 126 runs) was from ree-v2 synthetic scaffolding and is invalid. REE-native precision routing is addressed by MECH-059 directly. Retired. depends_on: - MECH-059 - ARC-005 - ARC-004 location: docs/architecture/agency_responsibility_flow.md#q-013 source: - docs/architecture/agency_responsibility_flow.md # Q-014 retired 2026-03-29 -- JEPA integration decoupled from REE. # The JEPA-framed question is moot, but the underlying insight (representation # invariances can hide ethically relevant distinctions in attribution) is # REE-native and transferred to MECH-137a (see below). - id: Q-014 title: "Do JEPA invariances hide ethically relevant distinctions in REE attribution contexts?" claim_type: open_question subject: invariance.ethical_relevance_blind_spot_risk polarity: asks status: legacy lifecycle_stage: retired adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-25T16:51:50.758372Z adjudication_decision_status: applied adjudication_recommendation: hybridize evidence_quality_note: > JEPA integration dropped from REE architecture 2026-03-29. The JEPA-specific framing is moot. The core insight -- that representation learning invariances can create attribution blind spots -- survives as a REE-native concern. Wang et al. (NeurIPS 2024) and Mehrabi et al. (ACM Surveys 2022) provide the relevant grounding. The REE-native formulation is: z_world encoder invariances that discard harm-predictive features create structural moral blind spots in E3 harm evaluation. This is addressed by MECH-128 (E1 goal conditioning) and the SD-008/SD-009 encoder design requirements. Retired as JEPA question; REE-native blind-spot concern tracked via evidence_quality_note on MECH-128 and SD-008. Prior synthetic evidence invalid. depends_on: - MECH-057 - MECH-059 - ARC-015 - ARC-004 location: docs/architecture/agency_responsibility_flow.md#q-014 source: - docs/architecture/agency_responsibility_flow.md - id: Q-015 title: "What is the smallest commit-boundary token that still supports reliable multi-timescale attribution?" claim_type: open_question subject: commitment.boundary_token_minimal_contract polarity: asks status: active lifecycle_stage: adjudicated adjudication_outcome: retain_ree adjudicated_at_utc: 2026-02-15T18:46:42.773429Z adjudication_decision_status: applied adjudication_recommendation: retain_ree evidence_quality_note: | Prior experimental evidence from claim_probe_mech_061 (8 runs) was from ree-v2 synthetic scaffolding (archived 2026-02-26) — those entries are invalid. Genuine ree-v1-minimal experiment completed 2026-02-26 (EVB-0041, commitment_boundary_validation, EXQ-003): PASS — MECH-061 confirmed. WITH-BOUNDARY harm < BLENDED, |corr|=0.114. This partially answers Q-015: the minimal contract requires at minimum that pre-commit and post-commit signals remain uncorrelated (|corr|<0.7) and that routing them separately outperforms blending. Both criteria confirmed at ree-v1-minimal scale. Demotion recommendation in promotion_demotion_recommendations.md (conflict_ratio=0.889, conf=0.55) is superseded — it was computed from the now-invalid synthetic entries. Retain active. Question is partially answered; further experiments on richer substrates will tighten the boundary contract definition. EXQ-155 SUPERSEDED (2026-03-29): implementation gap -- policy trained only with entropy bonus (ENT_BONUS=5e-3), z_world/z_self detached from policy forward pass. Policy converges to maximum-entropy random walk (action_entropy=ln(5)=1.6094); commit boundary never exercised. Redesign needed with harm-avoidance policy gradient wired to policy. See EXP-0094. depends_on: - MECH-061 - ARC-003 - ARC-015 location: docs/architecture/e3.md#q-015 source: - docs/architecture/e3.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - id: Q-016 title: "What arbitration policy best resolves tri-loop gate conflicts without coupling collapse?" claim_type: open_question subject: commitment.tri_loop_conflict_arbitration_policy polarity: asks status: active evidence_quality_note: | EXQ-156 SUPERSEDED (2026-03-29): implementation gap -- policy trained only with entropy bonus (ENT_BONUS=5e-3), z_world/z_self detached from policy forward pass. Policy converges to maximum-entropy random walk (action_entropy=ln(5)=1.6094); tri-loop gate conflicts never generated. Redesign needed with harm-avoidance policy gradient wired to policy. See EXP-0094. depends_on: - MECH-062 - ARC-003 - ARC-005 location: docs/architecture/e3.md#q-016 source: - docs/architecture/e3.md - docs/thoughts/2026-02-15_basal_ganglia.md - id: Q-017 title: "What is the minimal orthogonal control-axis subset that preserves observed regime separations?" claim_type: open_question subject: control_plane.minimal_orthogonal_axis_set polarity: asks status: active lifecycle_stage: adjudicated adjudication_outcome: hybridize adjudicated_at_utc: 2026-02-25T16:51:50.648196Z adjudication_decision_status: applied adjudication_recommendation: hybridize evidence_quality_note: | Prior experimental evidence from claim_probe_q_017 / control_axis_ablation (30 runs) was from ree-v2 synthetic scaffolding. Adjudication (hybridize) is valid architectural direction. Needs genuine ree-v1-minimal control axis experiments. EXQ-157 T081401Z SUPERSEDED (2026-03-29): dry_run=true -- no training occurred; all metrics zero; result invalid. Manifest marked superseded. EXQ-157 T185928Z SUPERSEDED (2026-03-29): implementation gap -- policy trained only with entropy bonus (ENT_BONUS=5e-3), z_world/z_self detached from policy forward pass. Policy converges to maximum-entropy random walk (action_entropy=ln(5)=1.6094); control axes never exercised. Redesign needed with harm-avoidance policy gradient wired to policy. See EXP-0094. EXQ-157 INCONCLUSIVE (2026-03-30, T065453Z): first valid run (prior runs superseded). PARTIAL_COLLAPSE_ADEQUATE: all three axis conditions (FULL/COLLAPSED/MINIMAL) fall within equivalence margin (C5_all_within_equiv=True). C4 (axes non-degenerate) FAIL -- axes likely collapsing in this substrate. Task complexity (8x8 grid) insufficient to discriminate axis structures; single-stream already saturates task capacity. Classified inconclusive: substrate inadequate to test Q-017, not evidence that minimal subset = 1 axis. Redesign needed at higher task complexity. depends_on: - MECH-063 - ARC-005 - MECH-055 location: docs/architecture/control_plane.md#q-017 source: - docs/architecture/control_plane.md - docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md - id: Q-018 title: "What RC-conflict threshold calibration blocks authority spoofing without chronic over-suppression?" claim_type: open_question subject: control_plane.reality_conflict_hysteresis_calibration polarity: asks status: active evidence_quality_note: | EXQ-158 PASS (2026-03-29): authority-spoofing event injection (2000+ events) confirmed RC-conflict gate fires correctly and spoofing suppression applies without chronic over- suppression. Note: experiment synthesizes spoofing events directly rather than relying on organic hazard encounters; result valid for the threshold calibration question but does not test the full organic harm-encounter pipeline. Partial answer. depends_on: - MECH-065 - MECH-062 - ARC-005 location: docs/architecture/control_plane_signal_map.md#q-018 source: - docs/architecture/control_plane_signal_map.md - docs/thoughts/17-02-26_necessary_separations_based_on_considering-prompt_injection.md - id: IMPL-020 title: "Formal JEPA/REE terminology alignment glossary for interoperability." claim_type: implementation_note subject: terminology.jepa_ree_alignment_glossary polarity: records status: stable depends_on: - IMPL-001 - MECH-057 location: docs/glossary.md#impl-020 source: - docs/glossary.md - docs/thoughts/2026-02-13_jepa_ree_formal_alignment_glossary.md - id: IMPL-021 title: "Hybrid JEPA/REE architecture diagram contract and rendering checklist." claim_type: implementation_note subject: diagrams.jepa_ree_hybrid_architecture_contract polarity: records status: stable depends_on: - IMPL-020 - MECH-057 - ARC-005 - ARC-015 location: docs/architecture/jepa_ree_hybrid_diagram_spec.md#impl-021 source: - docs/architecture/jepa_ree_hybrid_diagram_spec.md # JEPA flags removed 2026-02-26 — historical record only (JEPA decoupled). - id: IMPL-022 title: "JEPA-like E1/E2 representation-reference contract (inputs, outputs, knobs, failure gates)." claim_type: implementation_note subject: integration.jepa_like_e1e2_representation_reference_contract polarity: records status: stable lifecycle_stage: adjudication_enabled adjudication_mode: external_precedence_allowed allowed_conflict_outcomes: - retain_ree - hybridize - adopt_jepa_structure - retire_ree_claim depends_on: - IMPL-021 - MECH-057 - ARC-001 - ARC-002 - ARC-005 - ARC-015 location: docs/architecture/jepa_e1e2_integration_contract.md#impl-022 source: - docs/architecture/jepa_e1e2_integration_contract.md - id: IMPL-023 title: "REE-v2 representation-interface-first spec and phase gate." claim_type: implementation_note subject: roadmap.ree_v2_representation_interface_spec polarity: records status: candidate depends_on: - IMPL-008 - IMPL-021 - IMPL-022 - MECH-057 - MECH-058 - MECH-059 - MECH-060 - MECH-061 - MECH-062 - MECH-063 location: docs/architecture/ree_v2_spec.md#impl-023 source: - docs/architecture/ree_v2_spec.md - docs/roadmap.md - id: IMPL-024 title: "v2 REE-first canonical wording policy with JEPA interface translation." claim_type: implementation_note subject: terminology.v2_ree_first_canonical_jepa_interface_translation_policy polarity: records status: candidate depends_on: - IMPL-020 - IMPL-022 - IMPL-023 location: docs/notes/jepa_language_policy.md#impl-024 source: - docs/notes/jepa_language_policy.md - id: IMPL-025 title: "Cross-version hook surface contract and registry." claim_type: implementation_note subject: interfaces.cross_version_hook_surface_contract polarity: records status: candidate depends_on: - IMPL-022 - IMPL-023 - MECH-061 - MECH-062 - MECH-063 location: docs/architecture/hook_surface_contract.md#impl-025 source: - docs/architecture/hook_surface_contract.md - docs/architecture/hook_registry.v1.json - id: INV-019 title: "Rehearsal traversal and irreversible durable write must remain separated." claim_type: invariant subject: meta.selection_compression_boundary polarity: asserts status: candidate depends_on: - INV-011 - INV-004 - INV-006 - MECH-060 location: docs/invariants.md#inv-019 source: - docs/invariants.md - docs/thoughts/2026-02-18_selection_compression__symmetry_ and_meta_invariant_consolidation.md - id: INV-020 title: "Constraint stores are authority-stratified from direct observational/symbolic writes." claim_type: invariant subject: meta.authority_stratification_boundary polarity: asserts status: candidate depends_on: - INV-014 - INV-007 - MECH-064 - MECH-065 location: docs/invariants.md#inv-020 source: - docs/invariants.md - docs/architecture/control_plane_signal_map.md - docs/thoughts/2026-02-18_selection_compression__symmetry_ and_meta_invariant_consolidation.md - id: INV-021 title: "Responsibility-bearing durable updates occur only at typed commit boundaries." claim_type: invariant subject: meta.commit_boundary_irreversibility polarity: asserts status: candidate depends_on: - INV-012 - MECH-061 - Q-015 location: docs/invariants.md#inv-021 source: - docs/invariants.md - docs/architecture/e3.md - docs/thoughts/2026-02-18_selection_compression__symmetry_ and_meta_invariant_consolidation.md - docs/thoughts/2026-02-24_determinism_action_gating_boundary.md - docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md - id: INV-022 title: "Trust/precision allocation must remain heterogeneous, not a single scalar." claim_type: invariant subject: meta.heterogeneous_trust_allocation polarity: asserts status: candidate depends_on: - INV-008 - INV-009 - MECH-063 - Q-017 location: docs/invariants.md#inv-022 source: - docs/invariants.md - docs/architecture/control_plane_signal_map.md - docs/thoughts/2026-02-18_selection_compression__symmetry_ and_meta_invariant_consolidation.md - id: INV-023 title: "Protected offline recalibration/integration regimes are structurally required." claim_type: invariant subject: meta.offline_reweighting_requirement polarity: asserts status: candidate depends_on: - INV-010 - ARC-011 - MECH-016 - MECH-017 - MECH-018 location: docs/invariants.md#inv-023 source: - docs/invariants.md - docs/architecture/sleep.md - docs/thoughts/2026-02-18_selection_compression__symmetry_ and_meta_invariant_consolidation.md - id: INV-024 title: "Offline consolidation and online commitment must remain isolated at responsibility-bearing write loci." claim_type: invariant subject: meta.offline_online_update_locus_isolation polarity: asserts status: candidate depends_on: - INV-021 - INV-023 - ARC-020 - MECH-067 location: docs/invariants.md#inv-024 source: - docs/invariants.md - docs/architecture/sleep.md - docs/architecture/agency_responsibility_flow.md # Five foundational axioms — registered 2026-03-18 - id: INV-025 title: "Irreducible uncertainty is an epistemological constraint that cannot be engineered away." claim_type: invariant subject: epistemology.irreducible_uncertainty polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-025 source: - docs/architecture/five_axioms_foundations.md notes: > Axiom 1: "You can never be sure." Epistemic uncertainty is irreducible and structural — not a limitation to be overcome but a feature of being a finite agent in a real world. No agent can achieve certainty about its perceptions, predictions, or the causal consequences of its actions. Architectural consequence: the entire precision architecture (ARC-016, INV-008, INV-009) follows from this axiom. Pre-commit simulation (INV-011), commitment gates (INV-012), and precision-weighted prediction error are necessary responses to irreducible uncertainty. A fully certain agent needs no commitment boundary. - id: INV-026 title: "A self is an operational prerequisite for responsible agency." claim_type: invariant subject: self.existence polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-026 source: - docs/architecture/five_axioms_foundations.md notes: > Axiom 2: "I am." The agent exists as a distinct locus of experience, action, and responsibility. Without a self there is no causal agency, no commitment, no harm attribution, no accountability. The self is not an emergent computational convenience — it is a foundational axiom from which agency and ethics follow. Architectural consequence: z_self encoder (SD-005), the self/world split (z_self as E2 domain), the hypothesis tag that marks internally-generated content as distinct from world-received content (MECH-094), and the commitment boundary that makes actions attributable to an agent (INV-012). Also grounds the asymptotic unknowability of death: death is the limit of the harm gradient at which "I am" ceases — real (INV-027 makes it possible), but unknowable from within experience (ARC-024). - id: INV-027 title: "An external world is a structural prerequisite for grounded representation." claim_type: invariant subject: world.existence polarity: asserts status: active depends_on: [] location: docs/invariants.md#inv-027 source: - docs/architecture/five_axioms_foundations.md notes: > Axiom 3: "The world exists." There is a real world independent of the agent's model of it. The world can surprise the agent. World states are not constructions — they have causal power that predictions can get wrong. Architectural consequence: z_world encoder, prediction error channels (E1, E2, E3), the residue field (which accumulates what actually happened, not what was predicted), and post-commit trace persistence (INV-004, INV-006). Also grounds ARC-024: because the world exists independently, it can generate harm gradients that precede and predict harm events — the world is structured, not random, so proxy signals are real and informative. - id: INV-028 title: "Shared-world ethics requires modelling others as co-inhabitants of the same consequence space." claim_type: invariant subject: ethics.shared_world polarity: asserts status: active depends_on: - INV-026 - INV-027 location: docs/invariants.md#inv-028 source: - docs/architecture/five_axioms_foundations.md notes: > Axiom 4: "Others share this world." Other agents exist and inhabit the same world (INV-027). They are not simulations or projections of the self (INV-026). Their harm and benefit are real by exactly the same grounds as the agent's own — because they are selves (INV-026) in the same world (INV-027). Architectural consequence: harm via mirror modeling (INV-005) and ethical emergence from self-other symmetry (INV-001). Ethics is not an overlay because when another self is represented using the same predictive machinery as the self (INV-026), their harm generates the same error signal structure as own-harm. The self-other distinction is a routing difference within the same architecture, not a different architecture. - id: INV-029 title: "Love as long-horizon care-investment is a structurally coherent agent disposition, not an add-on sentiment." claim_type: invariant subject: ethics.love_exists polarity: asserts status: active depends_on: - INV-026 - INV-028 location: docs/invariants.md#inv-029 source: - docs/architecture/five_axioms_foundations.md notes: > Axiom 5: "Love exists." Genuine connection, care, and the pull toward union with others is a real phenomenon — not an overlay, not reducible to self-interest, not a computational strategy. Love is the asymptotic limit of the benefit gradient (ARC-024): every benefit signal (warmth, connection, belonging, joy) is a proxy pointing toward a complete union with another that is real (grounded in INV-028) but unknowable in its ultimate form. The unknowability is structural: complete union with another would approach the dissolution of the individuation that INV-026 asserts. Both selves cannot be fully present AND fully unified — the limit is asymptotically real but unreachable while both agents persist as distinct loci. Love can be approached in understanding and in experience; it cannot be fully known. Architectural consequence: grounds the benefit gradient direction (ARC-024), explains why harm-to-others matters (INV-028 applied to a being with a benefit manifold), and completes the ethical architecture. Ethics emerges from constraint (INV-015) applied to agents for whom love is real. # Proxy-gradient consequence of the five axioms — registered 2026-03-18 - id: ARC-024 title: "Harm and benefit signals have asymptotic proxy structure in world latent space." claim_type: architecture_hypothesis subject: world.harm_benefit_asymptotic_proxy_structure polarity: asserts status: provisional depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 location: docs/architecture/five_axioms_foundations.md evidence_quality_note: | EXQ-028 PASS (2026-03-18): Gradient dominance confirmed with random policy on CausalGridWorldV2. mean_dz_world_hazard_approach >> mean_dz_world_none. World generates observable signals before contact events. Direct confirmation that the simulated world must and does produce proxy-gradient fields. EXQ-029 PASS (2026-03-18): E3.harm_eval learned graded danger model on gradient world. none=0.373, approach=0.612, contact=0.666. 10× improvement over old world (EXQ-027). E3 extends harm model backward along causal chain from contact to approach — direct computational instantiation of the philosophical derivation. Both PASS runs support the architectural/philosophical layer of this claim (gradient world is necessary and learnable). EXQ-033 FAIL (2026-03-18): Tested the separable behavioral/capability-expansion prediction (approach detection grows faster than contact detection with training depth) which is now split out as ARC-026. The philosophical layer of ARC-024 is not implicated by EXQ-033 — C1-C3 all PASSED; only C4 (slope ratio) failed due to training instability (both signals peaked ep500 then degraded to ep1000, a depth-calibration artifact). Governance 2026-03-19: ARC-024 claim split. Philosophical/foundational layer retained as provisional (architecturally necessary, EXQ-028/029 support world-gradient structure). Behavioral capability-expansion prediction split to ARC-026 (candidate, v3). Conflict evidence in ARC-024 experiment pool reflects ARC-026's scope, not ARC-024 proper. notes: > Because the world exists (INV-027) and the self exists (INV-026), death is the asymptotic limit of the harm manifold — the world can end the self, but death cannot be experienced from within the agent's perspective because experiencing it would violate "I am" (INV-026). Every harm signal is a proxy gradient pointing toward this unknowable endpoint. Because love exists (INV-029) and others share this world (INV-028), unconditional love/union is the asymptotic limit of the benefit manifold — real but unknowable in its complete form while both selves persist. Architectural consequence for the simulated environment: the world MUST generate observable proxy-gradient fields around harm/benefit sources. Continuous, spatially distributed harm/benefit signals that precede and predict contact events are not a convenience — they are the correct model of how harm and benefit actually work. A world that generates binary harm signals only at contact cannot test E3.harm_eval gradient learning, cannot provide E2.world_forward with action-conditional training signal, and cannot support the SD-003 attribution pipeline. Implementation: CausalGridWorldV2 adds hazard_field and resource_field gradient arrays to the observation, with continuous harm signal proportional to proximity. See ree-v3/ree_core/environment/causal_grid_world.py. The further behavioral prediction — that approach detection grows faster than contact detection as a function of training depth — is a separable empirical claim registered as ARC-026. # Foundational definitions extracted from Philosophy repo — registered 2026-03-18 - id: INV-030 title: "Viability is defined relative to the agent's continuity in a shared world, not as a scalar reward." claim_type: invariant subject: cognition.viability_definition polarity: asserts status: active depends_on: - INV-025 - INV-027 location: docs/invariants.md#inv-030 source: - docs/architecture/five_axioms_foundations.md notes: > Cognition is defined as the maintenance of coherent input-output behaviour across time without catastrophic failure under permanent uncertainty. This definition is deliberately non-truth-seeking. Real cognitive systems — biological, artificial, institutional — operate under permanent uncertainty, partial observability, and delayed feedback. Under these conditions, cognition cannot be characterised as the acquisition of truth but as viability: the capacity to persist coherently without catastrophic failure. Follows from INV-025 (certainty is unavailable) and INV-027 (the world exists and can surprise the agent). A fully certain agent in a transparent world could in principle be truth-seeking; a finite agent in a real world cannot. Viability is what remains when truth-access is structurally denied. Architectural consequence: no REE component optimises toward ground truth. E1/E2 predictions are evaluated for temporal coherence and survivability under action, not correspondence to reality. E3 selects trajectories that preserve long-horizon viability, not accuracy. - id: INV-031 title: "Functional persistence in a shared world, not abstract truth, is the operational goal of cognition." claim_type: invariant subject: epistemology.functional_persistence_replaces_truth polarity: asserts status: active depends_on: - INV-025 - INV-027 - INV-030 location: docs/invariants.md#inv-031 source: - docs/architecture/five_axioms_foundations.md notes: > REE explicitly rejects privileged latent world states. Internal structures are not representations of reality — they are tools for viable engagement with it. Truth is replaced by functional persistence under intervention. A trajectory or internal model persists if and only if it demonstrates temporal survivability: internal coherence across time, compatibility with incoming signals, preservation of future predictability, and viability under interaction with other agents. Action serves as the primary epistemic probe. Perturbation — acting on the world and observing the outcome — reveals which trajectories remain coherent under intervention. This is the only epistemic test available to a finite agent under permanent uncertainty (INV-025) in a real world (INV-027) that can surprise it. Consequence: prediction errors may be tolerated when absorbing them preserves long-horizon viability. A framework that must minimise prediction error at every step cannot do this. REE replaces minimisation with constrained trajectory selection under the viability criterion (INV-030). - id: INV-032 title: "Moral agency requires both approach and avoidance drives; pure avoidance produces a degenerate risk manager, not an ethical agent." claim_type: invariant subject: agency.approach_avoidance_both_necessary polarity: asserts status: candidate depends_on: - INV-029 - INV-030 location: docs/invariants.md#inv-032 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > An architecture capable only of avoidance cannot care -- it can only refrain. Care (INV-029) requires orientation toward flourishing, not merely absence of harm. Both drives must be structurally represented. This is constitutive of ethical agency, not an optional architectural feature: a system that avoids all harm by doing nothing has satisfied every avoidance criterion while failing as an agent. Registered 2026-03-22. - id: INV-033 title: "REE agents require second-order epistemic access to their own model confidence, structurally represented and wired into commit gating, not just observable from performance metrics." claim_type: invariant subject: epistemic.second_order_self_monitoring polarity: asserts status: candidate depends_on: - INV-030 - INV-032 - ARC-016 - MECH-113 location: docs/architecture/approach_avoidance_symmetry.md#inv-033 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md notes: > REE's first-order error signals (E3 running_variance over z_world, E1 prediction error, E2 motor-sensory error) all track prediction accuracy about external or transitional states. None track the reliability of the self-model itself. Second-order access is the agent knowing how much to trust its own current z_self -- not just what z_self predicts, but whether z_self is in a state coherent enough to be trusted. This is architecturally necessary because: (1) world confidence and self-confidence are dissociable -- accurate world predictions coexist with an incoherent self-model; (2) commitment involves irreversible self-engagement, requiring reliable self-knowledge as a precondition; (3) the hypothesis tag (MECH-094) presupposes a coherent self that can distinguish simulation from real action -- when self-model is dispersed, the tag itself becomes unreliable (MECH-115). This is not an optimization detail: it is constitutive of the ethical architecture. Candidate mechanism: D_eff (participation ratio) monitoring of z_self (MECH-113/MECH-114). Hopfield stability (familiarity of z_self patterns vs stored states) is a complementary framing from epistemic-mapping repo (dgolden): stability = max softmax(beta * similarities) over 64-slot LRU memory. Combined certainty = 0.4*(1-entropy/10) + 0.3*(1-D_eff/hidden_size) + 0.3*stability provides a richer multi-framing implementation. The philosophical grounding is in Daniel's unpublished Synthese paper (Philosophy repo): ethical agency requires the agent to model its own epistemic state as a precondition for responsible action. Knowing one is not sure is not merely useful -- it is required for the hypothesis tag to be meaningful. Registered 2026-03-23. - id: ARC-025 title: "Three-engine (E1/E2/E3) architecture is irreducible; no engine can be collapsed into another." claim_type: architecture_hypothesis subject: architecture.three_engine_irreducibility polarity: asserts status: provisional depends_on: - INV-030 location: docs/architecture/five_axioms_foundations.md evidence_quality_note: | EXQ-034 PASS (2026-03-19): Direct irreducibility ablation. Three conditions (full, e3_ablated, e2_ablated) on CausalGridWorldV2, same warmup + eval. E3 ablation: cal_gap_approach collapses 0.192 → 0.000 (constant harm_eval=0.5). E2 ablation: attribution_gap collapses 0.034 → 0.000 (no counterfactual signal). Full system holds both simultaneously. Neither engine substitutes for the other. wf_r2 stable across conditions (0.947) confirming E2 world-forward unaffected by harm_eval ablation. PASS 5/5. Promoted active → provisional. notes: > Three functional constraints are necessary and sufficient for viable cognition. Neither two nor four constraints can satisfy all five axioms simultaneously. Two-constraint failures (any pair from E1/E2/E3): E1 + E2 without E3: coherent extrapolation without commitment. Produces analysis without action; temporal depth without decisiveness. Axiom III (self-persistence / viability) violated — the system cannot commit to futures that affect the self. E1 + E3 without E2: impulsive commitment without temporal depth. Produces reactive decisions without long-horizon integration. Axiom III violated again — consequences beyond the immediate step are not modelled. E2 + E3 without E1: brittle rigidity without responsiveness. Produces internally coherent plans that cannot update to rapid signal change. Axiom I (epistemic humility / you can never be sure) violated — the system cannot revise fast enough. Four-constraint redundancy: any decomposition of the triad produces functions already captured by E1/E2/E3. Additional constraints add complexity without new axiomatic coverage. The three-way tension between speed (E1), depth (E2), and commitment (E3) is irreducible. Architectural consequence: the three functional roles must remain distinct and in sustained interaction. Collapsing any two into one produces a recognisable failure mode. This is the architectural basis of the separation of E1/E2/E3 as distinct training targets with incommensurable error signals (MECH-069, ARC-021). - id: ARC-026 title: "Love as long-horizon care investment expands under intelligence rather than being crowded out." claim_type: architecture_hypothesis subject: ethics.love_expands_under_intelligence polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-024 - MECH-071 location: docs/architecture/five_axioms_foundations.md evidence_quality_note: | EXQ-033 FAIL (2026-03-18): Tested C4 -- approach_slope=0.000265, contact_slope=0.000289, ratio=0.920. FAIL attributed to training instability at ep500-1000 (depth-calibration artifact). C1-C3 all PASSED. EXQ-033 superseded by EXQ-232. EXQ-232 PASS 4/4 criteria, 3/3 seeds (2026-04-05): Peak-checkpoint analysis (1500 eps, phased E3 harm supervision from ep100). approach_slope consistently exceeds contact_slope at peak in all seeds. gap_at_peak=0.924-1.000. Resolves EXQ-033 depth-calibration artifact. Strong evidence: approach gradient expands faster than contact under training. notes: > Behavioral/capability-expansion prediction split from ARC-024 (2026-03-19). The philosophical/foundational layer (death/love as asymptotic limits) is ARC-024, which is well-supported and elevated to active. This claim contains the separable empirical prediction: under sufficient training depth, approach-gradient detection (the extended harm/benefit signal at distance) grows faster than contact detection. The reasoning is that contact detection is a simpler classification task — binary, high-signal — while approach detection requires the agent to learn a graded spatial manifold from noisier, lower-magnitude signals. Intelligence (training depth) should differentially benefit the harder learning problem. In functional terms: love (benefit gradient sensitivity, resource-approach detection) expands faster than raw harm-avoidance (contact avoidance) as the agent becomes more capable. This is an empirical prediction that requires careful training-depth calibration experiments to test cleanly. It is not a philosophical derivation — it could be false if contact detection saturates first due to magnitude advantage rather than approach detection accelerating. See ARC-024 for the foundational invariant structure from which this prediction is drawn. - id: MECH-102 title: "Violence is a terminal error-correction mechanism triggered only when all other channels fail." claim_type: mechanism_hypothesis subject: ethics.violence_as_terminal_error_correction polarity: asserts status: active evidence_quality_note: | EXQ-123 FAIL (2026-03-29): harm gradient pair FAIL. First experimental entry. depends_on: - INV-029 - INV-028 - INV-030 location: docs/architecture/five_axioms_foundations.md notes: > Violence is terminal error correction: maximal-energy intervention selected when low-energy coherent coordination pathways have been exhausted or are no longer available. This characterisation follows from INV-029 (love as the structural bias toward futures that preserve coordination). The default trajectory selection under love favours repair, exchange, and arbitration — the low-energy pathways that keep future coordination possible. Violence is not selected when low-energy pathways remain viable. Violence emerges when: (a) low-energy coordination pathways have been destroyed or are unavailable, (b) the agent's model projects that no coherent future is reachable without a high-energy corrective intervention, or (c) the commitment architecture has failed such that the agent cannot model consequences beyond the immediate step (ARC-025 two-constraint failure). This mechanism is scale-invariant. It applies at the level of individual behaviour, institutional action, and civilizational dynamics. In each case the structural signature is the same: escalation to violence is preceded by collapse of accessible low-energy coordination pathways, not by an absence of ethical constraint. Consequence for alignment: systems that destroy or narrow the space of available low-energy coordination pathways — through coercion, monopolisation of arbitration, or systematic elimination of repair options — predictably produce violence as a downstream consequence, regardless of the ethical framing applied to their actions. evidence_quality_note: | EXQ-032b PASS (2026-03-19): Energy escalation ladder confirmed via ttype split. Random policy + CausalGridWorldV2. causal_sig by state energy: none (safe locomotion): -0.032 hazard_approach (medium): +0.006 contact (high, agent+env combined): +0.017 Actions are progressively more consequential as the agent enters higher-energy interaction states. wf_r2=0.948. PASS 5/5. Note: EXQ-032 and EXQ-032c both FAIL — harm_exposure EMA operationalization does not work at harm_scale=0.02 (contact too rare for EMA to accumulate). The ttype-split operationalization (EXQ-032b) is the correct test for MECH-102. EXQ-036 (multi-step attribution) will test whether k=5 rollout sharpens the agent_caused vs env_caused contact-type distinction. EXQ-059c PASS (2026-03-21, 4/5): Contact rate reduction ~10× confirmed via SD-010 harm stream pipeline. Harm signal available during trajectory planning produces measurable approach-gradient avoidance, consistent with the violence-as-terminal- error-correction framing: maximum ethical leverage is in the approach gradient before contact, not at contact itself. Corroborates EXQ-032b energy ladder. NORMATIVE STATUS NOTE (2026-03-29): MECH-102 is a normative architectural claim about what a well-designed system should do -- not a description of how biological systems actually behave. Psychological literature (Anderson & Bushman 2002 GAM; Berkowitz 1989) shows humans frequently select violence before low-energy pathways are exhausted via hostile attribution bias and script activation -- these are MECH-102 failure modes, not its mechanism. The forward-model framing (violence selected when the agent's model projects no coherent future accessible via low-energy paths) is distinct from a temporal last-resort claim (try non-violent options first in sequence). The forward-model framing avoids Aloyo's (2015) just-war critique of temporal last-resort: pathway accessibility is a predictive judgment, not a required sequence of attempts. Open question: avoiding premature violence may require positive inhibitory mechanisms not present in human biology; equally, the claim permits violence as ethically required action when pathway closure is genuine. Both remain architecturally untested and warrant scrutiny under V4 ethical validation. # Three-gate BG gating architecture — registered 2026-02-27 - id: Q-019 title: "Three-gate vs single-gate BG architecture: are the three cortico-striatal loops anatomically distinct or convergent?" claim_type: open_question subject: basal_ganglia.three_gate_vs_single_gate_architecture polarity: open status: open evidence_quality_note: | Open architectural question arising 2026-02-27. Two competing models: (A) Single-gate: BG gate one action endpoint, evaluating three criteria simultaneously (sensorium readiness, thought/trajectory readiness, motor commitment). (B) Three-gate: BG implement three anatomically distinct gating loops — (1) Sensorium loop: limbic/beta-associated, gates what the system attends to — selecting not only from environmental input but from ongoing sensorimotor loops for E3 complex attention. No trajectory generation. Reticular thalamus / limbic BG substrate. Control plane E1/E2 update gates (precision/salience routing for world model updates) are SEPARATE mechanisms from this BG sensorium gate, operating on a different axis (model update eligibility vs attentional selection for thought). (2) Thought loop: associative/theta-associated, mediodorsal thalamus prevents bleed-back into sensorium. The E3 complex operates here. The hippocampus provides an affectively- weighted map of viable trajectories (shaped by the residue field) and proposes paths by navigating this terrain — not by running transition predictions. E2 provides fast local transition checks (cerebellum-like: state + action → next_state). E1 provides long-horizon world model context. E3 is a large complex: hippocampal map + trajectory proposal + tactical evaluation + commitment gating + back-projections suppressing candidates that fail criteria. DMN = thought loop free-running unconstrained (no sensorium input; hippocampal-mPFC episodic future projection in rest/prospective mode). (3) Action commitment loop: sensorimotor, E3 commit boundary. Gates motor execution. Automaticity gradient: if post-commit metrics improve, the task migrates to the sensorimotor loop — less sensorium BG gate overhead for expert performance. Harm residue re-elevates automated patterns for deliberate E3 oversight. NOTE ON E2: E2 is a cerebellum-like fast transition model (state + action → next_state), NOT a trajectory generator. Current ree-v1-minimal E2FastPredictor.generate_candidates() is performing hippocampal trajectory-proposal work under an incorrect label. Future refactoring will separate E2 (fast transition model, sensorimotor loop) from a dedicated hippocampal module (part of E3 complex, trajectory proposal and affective terrain map). Loops share thalamic substrate (mediodorsal and VL thalamus; reticular nucleus of thalamus as inter-loop routing switch) and computational elements (residue field spans all loops). Structural re-use maximised via thalamic gating. REE implications: current E3-only gating model conflates all three operations at action time. Three-gate model requires gating at three pipeline stages: before LatentStack.encode (sensorium loop), within E3 complex tactical evaluation using the hippocampal map (thought loop), and at E3 commit boundary (action loop). E3 as complex needs a feedback path suppressing thought-loop trajectory candidates — not present in ree-v1-minimal. MECH-057 scope revision pending — see MECH-057 evidence_quality_note. Relevant literature: O'Reilly & Frank 2006 (PBWM), Hazy/Frank/O'Reilly 2007, Aron et al. 2007 (STN hyperdirect), Brittain & Brown 2014 (beta oscillations), Buckner et al. 2008 (DMN), Crick 1984 / Zikopoulos & Barbas (reticular nucleus). depends_on: - MECH-057 - MECH-059 - MECH-061 source: - docs/thoughts/2026-02-27_three_gate_bg_architecture.md # Registered 2026-03-14 — three-loop learning channel separation (from thought 2026-03-14_three_bg_systems_error_signals.md) - id: ARC-021 title: "Three BG-like cortico-striatal loops require distinct learning channels." claim_type: architectural_commitment subject: basal_ganglia.three_loop_learning_channel_separation polarity: asserts status: provisional status_note: | Promoted candidate → provisional 2026-03-17. SD-003 experiment series (V3-EXQ-007–010) showed attribution mechanism produces correct-direction signal only when environment provides sufficient per-step variation — consistent with three distinct error channels requiring separate learning pathways. Dense hazards (EXQ-010, gap=+0.027) strongest signal. depends_on: - Q-019 - ARC-004 - MECH-069 - MECH-033 location: docs/architecture/three_loop_learning_channels.md#arc-021 source: - docs/thoughts/2026-03-14_three_bg_systems_error_signals.md - docs/architecture/three_loop_learning_channels.md - id: MECH-069 title: "Sensory prediction error, motor-sensory error, and harm/goal error are incommensurable and cannot be collapsed." claim_type: mechanism_hypothesis subject: latent_stack.three_loop_error_signal_incommensurability polarity: asserts status: stable status_note: | Promoted candidate → provisional 2026-03-17. V3-EXQ-009 (wider E2) showed capacity increase caused overfitting not better attribution — the E2 motor-sensory channel cannot learn harm attribution regardless of capacity. This is exactly what incommensurability predicts: harm signal must flow through a separate E3 channel, not be forced through E2. supersedes: MECH-058 supersedes_note: | MECH-058 framed the E1/E2 distinction as timescale separation (LR difference). MECH-069 identifies the deeper claim: the three loops operate on incommensurable error signals (sensory prediction / motor-sensory / harm+goal) that cannot be merged without misattributing credit. This is a functional-domain distinction, not a speed distinction. Cannot be properly tested until V3 z_self/z_world split (SD-005) is implemented. Anatomical grounding (added 2026-03-17): the three error channels are incommensurable partly because they correspond to three anatomically distinct white matter pathways (dorsal SLF / ventral occipitotemporal / lateral STS stream). See MECH-099 and evidence/literature/targeted_review_reafference_streams/ (Haak & Beckmann 2018). depends_on: - ARC-021 - MECH-033 - ARC-018 - MECH-099 location: docs/architecture/three_loop_learning_channels.md#mech-069 source: - docs/thoughts/2026-03-14_three_bg_systems_error_signals.md - docs/architecture/three_loop_learning_channels.md - id: MECH-070 title: "E2 is a conceptual-sensorium motor model with a planning horizon that exceeds E1." claim_type: mechanism_hypothesis subject: latent_stack.e2_conceptual_sensorium_motor_model polarity: asserts status: candidate depends_on: - ARC-021 - MECH-069 - MECH-033 - ARC-018 location: docs/architecture/three_loop_learning_channels.md#mech-070 source: - docs/thoughts/2026-03-14_three_bg_systems_error_signals.md - docs/architecture/three_loop_learning_channels.md evidence_quality_note: | The rollout_horizon > prediction_horizon ordering is valid for planning only if E1 co-evolves z_world during rollout (see MECH-135). Without co-evolution, a longer rollout worsens goal visibility -- E3 plans in a frozen world for N steps instead of a shorter frozen world. Correct neurological mapping: E1=cortical (slow LSTM, z_world domain, prediction_horizon=20); E2=cerebellar (fast efference-copy forward model, z_self domain, training horizon=1-step). E2 training horizon is separate from rollout_horizon and correctly short. See MECH-135 for the full parallel-rollout claim and EXQ-103/104/105 for experimental validation. EXQ-132 FAIL (2026-03-29): E2 motor model pair FAIL. First experimental entry for this specific claim. Training budget likely insufficient to discriminate E2 motor model horizon. EXQ-212 FAIL 2/4 (2026-04-03): E2 motor model discriminative pair FAIL. C1 PASS (E2 degrades faster than E1 at h=1: E2_r2_h1=-19 vs E1_r2_h1=-43 -- E2 actually better at single step). C4 PASS (n_windows=6986). C2 FAIL: E2_slope=-4051 vs E1_slope=-24.8 (E2 degrades MUCH faster than E1 across horizons -- opposite direction required for C2; C2 requires slope_e2 > slope_e1 confirming longer-horizon planning, but both R2 values are deeply negative indicating neither model can predict multi-step at V3 scale). C3 FAIL (r2_e2_h10=-19511 << threshold -1000 -- E2 completely fails at h=10). The MECH-070 claim that E2 has a longer planning horizon than E1 is not supported: E2 degrades much more steeply than E1. This may reflect that E2 is still primarily a 1-step transition model at V3 training budget. Hold at candidate. Second weakening entry. - id: MECH-071 title: "E2 harm prediction is better calibrated for agent-caused vs environment-caused transitions." claim_type: mechanism_hypothesis subject: e3.harm_eval_calibration_gradient_asymmetry polarity: asserts status: provisional implementation_phase: v3 depends_on: - MECH-070 - ARC-024 location: docs/architecture/sd_003_experiment_design.md#mech-071 source: - docs/architecture/sd_003_experiment_design.md evidence_quality_note: | EXQ-026 PASS (2026-03-18): E3.harm_eval calibration gap confirmed on CausalGridWorldV2 without proxy fields. calibration_gap=0.037 > 0.03 threshold. harm_pred_std=0.069. PASS 5/5 criteria. EXQ-029 PASS (2026-03-18): Much stronger confirmation on CausalGridWorldV2 with proxy gradient fields (ARC-024). calibration_gap_approach=0.239, calibration_gap_contact=0.209, reafference_r2=0.408. E3 learned graded danger model: none=0.373, approach=0.612, contact=0.666. Promoted from candidate -> provisional. v3_pending cleared. Subject corrected: was e2.harm_prediction_calibration_asymmetry -- E2 is NOT the harm evaluator. E3.harm_eval is the correct locus. EXQ-027 FAIL was the diagnostic that forced this architectural correction. EXQ-085 through EXQ-085d (2026-03-23/24): tagged MECH-071/INV-034 but ALL failed at goal-seeding bottleneck (z_goal_norm < 0.1 in every iteration). These do NOT contribute evidence for or against MECH-071's core claim (harm_eval calibration gradient). The mechanism under test (goal-state latent seeded by benefit exposure) was never instantiated because random-walk warmup without homeostatic drive cannot generate goal representations. Non-contributory: hold redesign pending drive/hunger architecture and curriculum schedule. See thought: docs/thoughts/2026-03-24_mech071_goal_latent_non_contributory_evidence.md EXQ-085e FAIL 3/4 (2026-03-26): z_goal_norm=0.135 > 0.1 threshold (C1 marginally PASS) but benefit_ratio=1.00 < 1.3 (C2 FAIL). C3 PASS (calibration_gap=0.1664). Root cause of C2=1.00 (wiring bug, 2026-03-27 diagnosis): In every 085x script, goal_state was maintained (seeding worked) but NEVER passed to action selection. Both GOAL_PRESENT and GOAL_ABSENT used random.randint() in eval -- a disconnected signal cannot change behaviour, guaranteeing benefit_ratio=1.00. E3Selector.score_trajectory() already has compute_goal_score() implemented (lines 453-457 of e3_selector.py): subtracts goal_weight * goal_proximity(z_world) from cost when goal_state.is_active(). The architecture is correct. The experiment scaffolding was broken across all 5 iterations. C3 consistently PASS across 085 through 085e (calibration_gap 0.03-0.17): this IS meaningful evidence for MECH-071's core claim (harm calibration intact with goal conditioning). Do not demote MECH-071 based on this series -- C2 failures are a wiring bug, not an architectural failure. EXQ-085f queued (2026-03-27): fixes the wiring. GOAL_PRESENT eval uses _goal_guided_action() -- E2.world_forward(z_world, a) for each candidate action, goal_state.goal_proximity() scores each prediction, greedy selection. GOAL_ABSENT still random. Same training for both conditions. New diagnostics: goal_resource_r (Pearson r: is z_goal pointing toward resources?) and goal_vs_harm_ratio (MECH-124 V4 risk: is z_goal salience competitive with harm?). EXQ-085f FAIL 3/4 (2026-03-27, 3-seed): C1 PASS (z_goal_norm=0.228 > 0.1). C3 PASS (cal_gap=0.030 -- MECH-071 calibration gradient intact). C4 PASS. C2 FAIL (benefit_ratio=0.28x < 1.3x: GOAL_PRESENT=0.180/ep vs GOAL_ABSENT=0.637/ep). Wiring now confirmed working; the new bottleneck is seeding quality: goal_resource_r=0.087 (< 0.2 threshold) -- z_goal is seeded but not pointing toward resources. MECH-124 diagnostic: goal_vs_harm_ratio=2.104 (> 0.3) -- salience is adequate, not the issue. C3 PASS confirms MECH-071 calibration gradient is intact even when goal guidance hurts navigation. Next step: warm-start z_goal initialization or stronger curriculum. EXQ-085g FAIL 3/4 (2026-03-29): contact-gated seeding (seeding z_goal exclusively at resource contact, benefit_exposure=1.0). C1 PASS (z_goal_norm=0.399 > 0.1). C3 PASS (cal_gap=0.218). C4 PASS. C2 FAIL (benefit_ratio=0.37x). MECH-124: goal_vs_harm=2.249 (no V4 risk). Root cause: goal_resource_r=0.066 < 0.2 -- z_world at contact encodes full scene, not resource specifically; features are not location-invariant across respawns. Dedicated z_resource encoder needed (SD-015). MECH-071 calibration claim intact (C3 consistently PASS); failure is in goal navigation, not harm calibration. notes: > After training, E3.harm_eval(z_world_t) is better calibrated before hazard_approach and agent_caused_hazard transitions than env_caused_hazard transitions or locomotion. E3 learns a graded danger model from z_world: approach to hazard increases harm_eval score continuously as proximity increases, not only at contact events. This is the gradient-sensing structure that ARC-024 predicts: the agent learns to detect harm gradients (approach → contact) rather than binary harm endpoints. Discriminability relies on z_world encoding proximity via hazard_field channels (CausalGridWorldV2) and SD-007 reafference correction removing self-motion contamination. - id: MECH-072 title: "Foreseeable-harm gating on residue accumulation reduces false attribution without degrading harm avoidance." claim_type: mechanism_hypothesis subject: residue.foreseeable_harm_gating_reduces_false_attribution polarity: asserts status: candidate v3_pending: true implementation_phase: v3 depends_on: - MECH-071 - MECH-060 location: docs/architecture/sd_003_experiment_design.md#mech-072 source: - docs/architecture/sd_003_experiment_design.md evidence_quality_note: | EXQ-028 FAIL (V2): Same root cause as EXQ-027. Foreseeable-harm gating depends on E2 discriminating agent-caused from env-caused harm — impossible without z_self/z_world split (SD-005) and joint SD-003 pipeline. V3 EXQ-054 FAIL (2026-03-20): world_delta_agent=0.01776 ≈ world_delta_env=0.01806 (1.7% difference, below C3 threshold). C4 FAIL: calibration_gap_approach=0. E2 cannot discriminate agent vs environment dynamics without SD-005 z_self/z_world split — z_gamma conflates both. Partial signal: false_attr dropped 33%→26%, showing the gate threshold captures some structure, but C3/C4 require SD-005 to pass. v3_pending kept (2026-03-28): specific blocker is SD-011 (z_harm_s stream). Foreseeable-harm discrimination requires E2_harm_s forward model (ARC-033) which depends on SD-011 dual nociceptive split. New architectural question (2026-03-20): world-delta magnitude is insufficient to discriminate agent vs environment causation -- both produce similar E2 perturbation magnitudes. The gate requires directional discrimination: which subspace of z_world was perturbed, not just how much. Candidate fix: replace scalar ||delta|| threshold with a learned discriminator head trained on (z_world_delta, is_agent_caused) labels, using the oracle ground truth during training. This transforms MECH-072 from a threshold gate into a supervised discriminator gate. Requires SD-005 z_self/z_world split as prerequisite (C3 failure is fundamentally about E2 conflating agent and environmental dynamics in the same z_world space). Governance note (2026-03-22): governance algorithm produced a false promote_to_provisional recommendation (conf=0.951) despite 0 genuine supporting experiments. Root cause: consistent-weakening evidence (0 supports, 1 weakens, 4 mixed) produced high consistency score in the old formula. Both algorithm bugs fixed in build_experiment_indexes.py (directional consistency formula + min_supporting_entries gate). v3_pending restored. Hold at candidate until SD-005 + SD-010 implemented. Literature pulls for the discriminator gate design and biological evidence for causation-attribution gating mechanisms needed -- see evidence backlog. EXQ-213 PASS 6/6 (2026-04-03): Foreseeable-harm gating discriminative pair PASS. Clean PASS on all 6 criteria across 3 seeds. false_attr_rate gated=0.0 vs ungated=0.478 (48% reduction). harm_rate preserved (0.873 in both). e2_world_r2=0.558 (above C6 threshold). V3 proxy test: E2.world_forward predicts foreseeable harm; GATED accumulates residue only when E2 harm_eval(z_w_pred) > 0.2. This confirms the gating mechanism works at V3 scale using the existing E2 world model (not E2_harm_s -- using SD-010 z_harm pathway). Significant positive evidence for MECH-072 core claim. v3_pending kept: E2_harm_s (SD-011, ARC-033) is still needed for the full discriminator-gate design, but the V3 proxy is valid. notes: > Gating residue accumulation on E2 harm foreseeability (predict_harm > threshold) reduces false attribution rate (residue at env-caused events / total residue) versus naive accumulation at all harm events, without degrading harm avoidance. EXQ-028 tests this. # Registered 2026-03-15 — NC-01 through NC-09 valenced hippocampal map cluster # ⚠ ARC-022 and MECH-073 conflict with ARC-007 (hippocampal valence constraint) — pending adjudication via Q-020 - id: ARC-022 title: "Hierarchical effect-object abstraction pipeline (E1→E2→DMN→Goal/Avoid→Hippocampus) — conflict with ARC-007/ARC-014, pending adjudication." claim_type: architectural_commitment subject: architecture.hierarchical_effect_object_pipeline polarity: asserts status: provisional conflict_with: - ARC-007 - ARC-014 conflict_note: | ARC-007 constrains hippocampus to "no value computation"; ARC-022 calls it a "valenced cognitive map" at the pipeline's terminus. ARC-014 defines DMN as an operating mode; ARC-022 treats it as a pipeline layer. Resolution options: (A) revise ARC-007 to distinguish "compute new value" (denied) from "embody geometrically-encoded prior value" (permitted); (B) reject MECH-073 and treat NC-01 as external neuroscience model only. Pending Q-020 adjudication. depends_on: - ARC-001 - ARC-002 - ARC-007 - ARC-014 - ARC-021 - MECH-070 location: docs/architecture/valenced_hippocampal_map.md#arc-022 source: - docs/architecture/valenced_hippocampal_map.md - id: MECH-073 title: "Valence is intrinsic to hippocampal map geometry, not applied downstream — conflict with ARC-007, pending adjudication." claim_type: mechanism_hypothesis subject: hippocampal.intrinsic_map_valence polarity: asserts status: candidate evidence_quality_note: > Governance 2026-04-03: hold_candidate_resolve_conflict applied. conf=0.689, conflict_ratio=1, 0 exp entries, 5 lit entries. Literature-only conflict: 2 supports, 2 weakens, 1 mixed across lit entries. No experimental evidence yet. Conflict with ARC-007 noted below. Hold at candidate pending experimental discrimination. conflict_with: - ARC-007 conflict_note: | ARC-007 design constraint: "No value computation. Its function is orthogonal to valuation and control." MECH-073 asserts valence is intrinsic to hippocampal map geometry and that rollouts are inherently valenced simulations. Direct contradiction. Pending Q-020. depends_on: - ARC-007 - ARC-013 - MECH-035 location: docs/architecture/valenced_hippocampal_map.md#mech-073 source: - docs/architecture/valenced_hippocampal_map.md - id: MECH-074 title: "Amygdala functions as read/write head for valenced hippocampal map." claim_type: mechanism_hypothesis subject: amygdala.hippocampal_map_readwrite_head polarity: asserts status: candidate depends_on: - MECH-046 - MECH-073 - ARC-007 - ARC-013 location: docs/architecture/valenced_hippocampal_map.md#mech-074 source: - docs/architecture/valenced_hippocampal_map.md notes: > Extends MECH-046 (amygdala as mode prior updater) with four read/write roles on the hippocampal map: (a) encoding modulation via BLA projections; (b) retrieval biasing toward threat-relevant regions under arousal; (c) rapid pre-cortical priming via fast subcortical route; (d) triggered remapping on prediction error spike. The fast subcortical priming role is not present in MECH-046. - id: MECH-075 title: "Basal ganglia perform dopaminergic gain/threshold setting on hippocampal attractor dynamics." claim_type: mechanism_hypothesis subject: basal_ganglia.dopaminergic_attractor_threshold_setting polarity: asserts status: candidate depends_on: - Q-019 - ARC-021 - MECH-043 - MECH-073 location: docs/architecture/valenced_hippocampal_map.md#mech-075 source: - docs/architecture/valenced_hippocampal_map.md notes: > BG perform gain/threshold setting on attractor dynamics in the valenced hippocampal system via dopaminergic tone (VTA/SNc) and striatal competitive inhibition, not external value comparison. Extends MECH-043 (dopamine-like precision-weighting) with attractor- basin threshold framing specific to the hippocampal planning context. Supports three-loop architecture of ARC-021 with a mechanism for how each loop's gate operates. ANATOMICAL SUBDIVISION NOTE (2026-04-02, Kempadoo et al. 2016 PNAS): VTA dopamine axons are sparse in dorsal hippocampus. Dopamine in dorsal HPC (spatial/planning substrate, E3 viability map) comes primarily from locus coeruleus (LC) via DA/NE co-release, responding to novelty and arousal. VTA dopamine primarily targets ventral HPC (motivational/affective). MECH-075 bifurcates by HPC axis: dorsal=LC-mediated arousal gain; ventral=VTA-mediated RPE/reinforcement gain. anatomical_note: "LC (locus coeruleus), not VTA, is primary catecholamine source in dorsal HPC (spatial/planning terrain). Dorsal HPC gain: LC-mediated (novelty/arousal, DA/NE co-release, Kempadoo et al. 2016). Ventral HPC gain: VTA-mediated (reward prediction error, RPE signal). MECH-075 survives but requires axis-specific implementation: dorsal terrain encoding-depth modulation via LC arousal signal; ventral valence weighting via VTA RPE." evidence_quality_note: | EXQ-192a FAIL 1/4 x2 runs (2026-04-03): Hippocampal-VTA novelty loop probe FAIL. Run 1 (T04:39): criteria_met=1/4. Run 2 (T10:25): criteria_met=1/4. mean_novelty_signal_on = 6.4e-05 (threshold C4 > 1e-04: FAIL). cell_gap=0, hazard_gap=0 across both conditions -- NOVELTY_LOOP_ON identical to NOVELTY_LOOP_OFF. Root cause: novelty_gain=2.0 but novelty signal itself is below detection threshold; CEM noise scaling has no measurable effect when the novelty_ema is too small. The manipulation requires the novelty signal to exceed 1e-04 before behavioral differentiation is visible. Substrate limitation: E1 world-prediction error may be too uniform across z_world states at V3 training scale, or the CEM noise scaling (novelty_gain=2.0) is insufficient. C3 PASS (harm_delta=0 both conditions -- harm not worsened). Consistent null result across two independent runs. Hold at candidate; not falsified -- the novelty signal magnitude is the bottleneck, not the claim's core mechanism. Mixed/null evidence partially reflects dorsal/ventral axis split (Kempadoo et al. 2016): EXQ-192a probed VTA-like novelty gain on dorsal HPC terrain, but dorsal HPC gain is LC-mediated, not VTA-mediated. The FAIL is consistent with targeting the wrong catecholamine source for dorsal terrain. This is not a falsification of MECH-075 core mechanism -- it is a mis-targeting of anatomical substrate. See anatomical_note. - id: MECH-076 title: "Residue is structural deformation of hippocampal map topology (CA3 retraction, DG neurogenesis, reconsolidation)." claim_type: mechanism_hypothesis subject: residue.hippocampal_map_structural_deformation polarity: asserts status: candidate depends_on: - ARC-013 - MECH-056 - MECH-073 location: docs/architecture/valenced_hippocampal_map.md#mech-076 source: - docs/architecture/valenced_hippocampal_map.md notes: > Residue (ARC-013 curvature field) is instantiated as structural deformation of hippocampal map topology: CA3 dendritic retraction under stress, reduced DG neurogenesis, palimpsest traces from incomplete reconsolidation. Not a separate inhibitory layer — the terrain itself is deformed. Compatible with MECH-056 (trajectory-first placement): basin distortion acts as trajectory-level pressure without requiring a representational filter layered on perception. - id: MECH-077 title: "Therapeutic change equals hippocampal remapping, with modalities operating at different levels." claim_type: mechanism_hypothesis subject: hippocampal.therapeutic_remapping polarity: asserts status: candidate depends_on: - ARC-013 - MECH-076 - MECH-018 - MECH-073 location: docs/architecture/valenced_hippocampal_map.md#mech-077 source: - docs/architecture/valenced_hippocampal_map.md notes: > New domain claim. Therapeutic change = hippocampal remapping. Modalities: CBT operates on OFC/vmPFC read of existing map; EMDR operates via reconsolidation under low arousal (REE analogue: offline consolidation with arousal suppressed, MECH-018); psychodynamic therapy = slow relational geometry restructuring over many consolidation cycles; psychedelic- assisted = transient attractor flattening enabling large-scale map restructuring. - id: MECH-078 title: "Amygdala bootstraps novel valence for unmapped hippocampal territory; anxiety disorders are systematic over-valencing." claim_type: mechanism_hypothesis subject: amygdala.novel_territory_valence_bootstrap polarity: asserts status: candidate depends_on: - MECH-046 - MECH-074 - MECH-073 location: docs/architecture/valenced_hippocampal_map.md#mech-078 source: - docs/architecture/valenced_hippocampal_map.md notes: > Amygdala provides bootstrap valence for unmapped hippocampal territory via fast subcortical priming (MECH-074 role 3), then writes initial valence into map through consolidation. Anxiety disorders may reflect systematic over-valencing of novel/unmapped territory as aversive. Positive feedback loop with MECH-040 volatility channel. Extends MECH-046. - id: MECH-079 title: "Phenomenological continuous selfhood is an artefact of stable hippocampal map geometry." claim_type: mechanism_hypothesis subject: hippocampal.map_geometry_selfhood polarity: asserts status: candidate depends_on: - ARC-007 - ARC-013 - MECH-024 - MECH-073 location: docs/architecture/valenced_hippocampal_map.md#mech-079 source: - docs/architecture/valenced_hippocampal_map.md notes: > Phenomenological continuous selfhood is an artefact of stable hippocampal map geometry: same topology → similar rollouts → similar commitments → subjective identity persistence. Extends MECH-024 (selfhood/personality/ethics converge structurally) with specific substrate. Identity-level change = phase transition in map geometry (discontinuous restructuring of attractor basin profile). Co-constituted with INV-006: ethical history is structurally encoded in map geometry. - id: MECH-080 title: "Rollout truncation set-points as psychiatric individual differences substrate (ADHD/anxiety/OCD)." claim_type: mechanism_hypothesis subject: hippocampal.rollout_truncation_psychiatric_individual_differences polarity: asserts status: candidate depends_on: - ARC-018 - MECH-027 - ARC-014 - MECH-075 location: docs/architecture/valenced_hippocampal_map.md#mech-080 source: - docs/architecture/valenced_hippocampal_map.md notes: > Rollout sampling depth and truncation set-points as substrate for psychiatric individual differences. ADHD = truncation-biased (BG urgency threshold too low, premature commitment); anxiety = extension-biased (uncertainty chronically elevated, BG threshold too high, suppressed commitment); OCD = attractor lock-in (abnormally deep basin, MECH-076, normal signals cannot trigger basin exit). Extends MECH-027 (pathological modes = mis-tuned control-plane regimes) and ARC-014 failure modes with unified rollout-parameter account. - id: Q-020 title: "Does ARC-007's no-value-computation constraint survive MECH-073 (valence intrinsic to hippocampal map)?" claim_type: open_question subject: hippocampal.arc007_valence_constraint_survival polarity: open status: resolved resolution: resolution_a resolution_note: | RESOLVED 2026-04-02 via literature synthesis. Resolution A adopted: ARC-007 and MECH-073 are co-true. Hippocampus embodies value-shaped geometry written by external systems (BLA via BTSP plateau potentials; VTA dopamine; LC catecholamines) but does not compute value itself. The "no value computation" constraint specifies that hippocampus does not evaluate RPE, utility, or outcomes -- it stores the geometrically encoded result of computations performed externally. Key papers: Bittner 2017 (BTSP mechanism -- external write triggers), Teyler & Rudy 2007 (indexing theory -- HPC as pure relational binder), McGaugh 2004 + Dolcos 2004 (BLA modulates write depth). Gauthier 2018 dedicated reward cells reinterpreted as stable BLA-tagged anatomical relay stations, not value-computing units. See hippocampal_literature_synthesis_2026.md Section 1 for full ruling. conflict_note: | Pivotal adjudication question for NC-01 through NC-09 cluster integration into canonical REE. Two resolution paths: (A) Revise ARC-007 to distinguish "hippocampus does not compute new value" (still denied) from "hippocampus embodies geometrically-encoded prior value via amygdala write operations and residue-field shaping" (permitted under MECH-073). Preserves INV-014. (B) Reject MECH-073; treat NC-01/NC-02 as external neuroscience model that inspires but does not enter canonical architecture. ARC-022, MECH-073 demoted to rejected. MECH-074 through MECH-080 survive independently (most do not require MECH-073 to be valid). Note: MECH-074 through MECH-080 have partial independence from MECH-073 -- see individual conflict notes. Q-019 resolution evidence may inform this question. EXQ-159 SUPERSEDED (2026-03-29): implementation gap -- policy trained only with entropy bonus, z_world/z_self detached from policy forward pass; maximum-entropy random walk (action_entropy=ln(5)=1.6094). Hippocampal valence constraint (ARC-007) never exercised. Gated on SD-004 hippocampal navigation implementation. See EXP-0095. depends_on: - ARC-007 - MECH-073 - ARC-022 - Q-019 location: docs/architecture/valenced_hippocampal_map.md#q-020 source: - docs/architecture/valenced_hippocampal_map.md # Registered 2026-03-22 -- approach/avoidance drive symmetry: open question - id: Q-021 title: "Does training with harm-avoidance signals only, without explicit approach drives, produce behavioral flatness (quiescent degenerate policy)?" claim_type: open_question subject: drive.behavioral_flatness_under_pure_avoidance polarity: open status: open evidence_quality_note: | narrow_open_question (2026-03-29): two independent pathways to behavioral flatness already identified in notes. Pathway A (drive absence, ARC-030) tested by EXQ-072; Pathway B (self-incoherence-gated commit suppression, MECH-113/114) untested. Definitive test requires both pathways. No experimental evidence yet -- question narrowing applied to focus on discriminating between the two pathways. depends_on: - ARC-030 - MECH-111 - MECH-112 - MECH-113 location: docs/architecture/approach_avoidance_symmetry.md#q-021 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > Testable prediction of ARC-030. An ablation experiment trains two agents identically except one has approach drives (goal attractor, novelty signal, self-maintenance) and one does not. If ARC-030 is correct, the avoidance-only agent converges to minimal-action behavior measurable as reduced policy entropy and action rate over training. Secondary question: what minimum set of approach signals is sufficient to prevent behavioral collapse? Added to evidence backlog as EVB-0061. Registered 2026-03-22. TWO INDEPENDENT PATHWAYS TO BEHAVIORAL FLATNESS (registered 2026-03-23): Pathway A (drive absence -- ARC-030 route): No Go drive -> only avoidance gradient -> quiescent policy is optimal -> behavioral flatness. Tested by EXQ-072. Pathway B (self-incoherence-gated commit suppression -- MECH-113/114 route): Go and NoGo drives both present, but high D_eff -> MECH-114 suppresses commitment -> agent cannot act even when approach drive is active -> behavioral flatness through self-model failure. These pathways are independent and dissociable: an agent can PASS EXQ-072 (Pathway A resolved) but still exhibit flatness under Pathway B if self-model is disrupted. A definitive test of Q-021 requires testing both pathways. Pathway B experiment not yet written; will follow EXQ-075 (MECH-113) if PASS. The two pathways predict different signatures: Pathway A flatness is present even at low D_eff (coherent agent, just no goal drive); Pathway B flatness occurs despite Go drive being present, colocated with high D_eff. Clinical analogues (from literature): Schizophrenia negative symptoms (avolition) = impaired wanting with intact liking -- maps to Pathway A. Dissociative states where action is suppressed despite motivation -- maps more to Pathway B. Barch & Dowd 2010 (DOI 10.1093/schbul/sbq068), Culbreth et al. 2023 (DOI 10.1017/S0033291722003993). Depression = locked-in avoidant brain; allostatic self-efficacy failure -> learned helplessness; Stephan et al. 2016 (DOI 10.3389/fnhum.2016.00550), Barrett et al. 2016 (DOI 10.1098/rstb.2016.0011). Full synthesis: evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md # Registered 2026-03-15 — NC-02 through NC-03 E1/E2 constraint propagation cluster - id: MECH-081 title: "E2 sufficiency constraint reduces E1 effective dimensionality target." claim_type: mechanism_hypothesis subject: e1_e2.e2_sufficiency_constraint_reduces_e1_dimensionality polarity: asserts status: candidate depends_on: - ARC-001 - ARC-002 - MECH-033 location: docs/architecture/e1_e2_constraint_propagation.md#mech-081 source: - docs/architecture/e1_e2_constraint_propagation.md notes: > E2's prediction task generates a top-down sufficiency constraint on what E1 must resolve. E2 implicitly specifies which covariances in the sensory stream E1 must track for E2's predictions to be well-formed, reducing E1's effective dimensionality target. The hard problem of feature discovery is partially specified from above rather than requiring purely unsupervised discovery from below. May reduce E1 hidden-layer depth requirement. Extends ARC-001/ARC-002/MECH-033: E2 is both a consumer of E1 representations and an active shaper of what E1 must produce. - id: MECH-082 title: "Hippocampal map distortion propagates through E2 to bias E1 attentional sampling without explicit E1 retraining." claim_type: mechanism_hypothesis subject: e1_e2.top_down_perceptual_bias_propagation polarity: asserts status: candidate depends_on: - MECH-081 - MECH-076 - MECH-027 - ARC-001 - ARC-002 location: docs/architecture/e1_e2_constraint_propagation.md#mech-082 source: - docs/architecture/e1_e2_constraint_propagation.md notes: > Changes in E2's predictive model (shaped by hippocampal map distortion, MECH-076) propagate downward as changes in E1 perceptual sampling without explicit E1 retraining. Map distortion → biased E2 predictions → E2 sufficiency constraint (MECH-081) directs E1 to over-sample residue-consistent sensory dimensions. Provides a computational account of perceptual bias in trauma (PTSD hypervigilance), paranoia, and depression. Compatible with MECH-056 (trajectory-first residue): mechanism operates via trajectory-level basin geometry, not direct E1 representational distortion. # Registered 2026-03-15 — NC-11 through NC-16 four-plane neuromodulatory control cluster # ⚠ MECH-085 conflicts with MECH-006 (serotonin temporal depth prohibition) — pending Daniel's adjudication - id: MECH-083 title: "Acetylcholine as meta-level plasticity gain governing durable-write vs read-through." claim_type: mechanism_hypothesis subject: acetylcholine.meta_level_plasticity_gain polarity: asserts status: candidate depends_on: - ARC-005 - ARC-007 - ARC-018 - INV-019 - INV-024 location: docs/architecture/neuromodulatory_control_planes.md#mech-083 source: - docs/architecture/neuromodulatory_control_planes.md notes: > ACh is the meta-level control plane: modulates plasticity gain on bottom-up input — how strongly current sensory experience modifies the hippocampal map (high ACh) vs. how strongly existing attractors dominate interpretation (low ACh). NOT a binary switch but a continuous gain parameter. Distinct from NA: ACh governs whether attended input leaves a lasting map trace (learning timescale); NA governs whether the system orients to bottom-up surprise right now (moment-to-moment). Meta-level because it determines whether other planes' effects consolidate durably. Failure: OCD-relevant (high ACh during retrieval deepens stuck attractors); chronic high = map instability; chronic low = learning failure / confirmation bias. Functionally instantiates INV-019/INV-024 (rehearsal vs durable write separation) at hippocampal level. - id: MECH-084 title: "Noradrenaline as attentional snap and E1/E2 sampling ratio modulator." claim_type: mechanism_hypothesis subject: noradrenaline.attentional_snap_and_e1_e2_sampling_ratio polarity: asserts status: candidate depends_on: - ARC-001 - ARC-002 - MECH-005 - MECH-081 - ARC-005 location: docs/architecture/neuromodulatory_control_planes.md#mech-084 source: - docs/architecture/neuromodulatory_control_planes.md notes: > NA provides the attentional snap: fast moment-to-moment orienting to bottom-up surprise. High NA reduces E2's constraining influence on E1 (MECH-081), snapping current attention to unexpected sensory signals. Distinct from ACh (MECH-083): NA = does the system orient to this input now? ACh = does that orientation leave a lasting map trace? Together: high NA + high ACh = maximum new learning from unexpected input; high NA + low ACh = aroused/vigilant but not learning; low NA + high ACh = steady ambient updating. Too little: perceptual confirmation bias (E2 dominates). Too much tonic: pervasive snap to noise (PTSD hypervigilance). Too much phasic/dysregulated: episodic E1/E2 decoupling (candidate dissociation mechanism). Extends MECH-005 with upstream E1/E2 interface account. - id: MECH-085 title: "Serotonin as hippocampal map geometry parameter (rollout depth, distal value weighting) — conflict with MECH-006, pending adjudication." claim_type: mechanism_hypothesis subject: serotonin.hippocampal_map_geometry_parameter polarity: asserts status: candidate conflict_with: - MECH-006 conflict_note: | MECH-006 explicitly states serotonin "does NOT select temporal depth (τ)." MECH-085 claims serotonin controls temporal depth of rollouts and modulates aversive attractor basin depth. Candidate reconciliation (raphe/receptor-subtype framing): multiple raphe nuclei and receptor subtypes mean serotonin may re-use the same motif (modulating resolution pressure/exclusivity) at different hierarchical levels. MECH-006 (dorsal raphe → cortical): cognitive-narrative collapse/exclusivity. MECH-085 (median raphe → hippocampal CA1/DG): same motif as attractor basin depth and rollout depth. If accepted, MECH-006's "no temporal depth τ" holds for the E-stack level; MECH-085's rollout depth is a distinct hippocampal-terrain construct. Literature pull required (raphe projection specificity, receptor-subtype distribution) before reconciliation can be accepted. Pending adjudication. depends_on: - MECH-006 - MECH-073 - MECH-076 - ARC-013 location: docs/architecture/neuromodulatory_control_planes.md#mech-085 source: - docs/architecture/neuromodulatory_control_planes.md notes: > Serotonin as map geometry parameter: (1) temporal depth of rollouts — how far forward simulation extends; (2) distal/proximal value weighting — high 5-HT flattens temporal distance penalty on far attractors; (3) aversive attractor basin depth — modulates steepness of threat-related attractor walls, explaining SSRI anxiolytic effects and 2-4 week onset delay. Extends MECH-006 (representational collapse/exclusivity) and MECH-076 (structural map deformation). Serotonin is a map geometry parameter, not a mood neuromodulator. Conflict with MECH-006 requires adjudication before promotion beyond candidate. - id: MECH-086 title: "Dopamine as trajectory selection gain plane operating downstream of landscape, state, and encoding gate." claim_type: mechanism_hypothesis subject: dopamine.trajectory_selection_gain_plane polarity: asserts status: candidate depends_on: - MECH-043 - MECH-075 - MECH-085 - ARC-005 location: docs/architecture/neuromodulatory_control_planes.md#mech-086 source: - docs/architecture/neuromodulatory_control_planes.md notes: > Dopamine sets gain on trajectory selection — which rollout wins and how quickly — by modulating attractor dynamics in the already-valenced hippocampal system. Exploitation signal at planning level. Key claim: dopamine optimises selection over whatever landscape (serotonin, MECH-085) and state representation (NA, MECH-084) and encoding mode (ACh, MECH-083) it receives — cannot compensate for upstream plane failures. Too little: indecision, flattened landscape, anhedonia. Too much: premature commitment, aberrant salience. Extends MECH-043 (unsigned PE precision-weighting) and MECH-075 (BG attractor threshold) with the hierarchical-constraint claim and explicit four-plane positioning. - id: MECH-087 title: "Hierarchical ordering of four neuromodulatory control planes: ACh → NA → 5-HT → DA." claim_type: mechanism_hypothesis subject: neuromodulation.four_plane_hierarchical_ordering polarity: asserts status: candidate depends_on: - MECH-083 - MECH-084 - MECH-085 - MECH-086 - ARC-005 - MECH-063 location: docs/architecture/neuromodulatory_control_planes.md#mech-087 source: - docs/architecture/neuromodulatory_control_planes.md notes: > Four neuromodulatory planes are hierarchically ordered by upstream position: ACh (meta-gate) → NA (perceptual sampling) → 5-HT (map geometry) → DA (selection gain). Each downstream plane operates over outputs of all upstream planes. Predictions: (1) cascade failure — dysfunction at any level cascades down but produces qualitatively distinct failure modes; (2) intervention order matters — upstream planes should be stabilised before downstream optimisation; (3) dopamine augmentation on distorted serotonergic landscape worsens outcomes. Extends MECH-063 (orthogonal control axes) with hierarchical asymmetry: axes are ordered, not merely orthogonal. Testable: degrading 5-HT axis produces trajectory pathology not rescued by DA augmentation. - id: MECH-088 title: "Psychiatric conditions as four-plane neuromodulatory control failures." claim_type: mechanism_hypothesis subject: neuromodulation.psychiatric_conditions_four_plane_failures polarity: asserts status: candidate depends_on: - MECH-083 - MECH-084 - MECH-085 - MECH-086 - MECH-087 - MECH-080 - MECH-027 location: docs/architecture/neuromodulatory_control_planes.md#mech-088 source: - docs/architecture/neuromodulatory_control_planes.md notes: > Psychiatric conditions as four-plane control failures: PTSD = 5-HT map geometry failure (deep aversive attractors) + secondary NA hyperactivity; Depression = 5-HT distal attractor inaccessibility + DA selection failure; ADHD = NA inconsistent sampling + DA unstable attractor commitment; OCD = DA selection stuck in local attractor + ACh gating failure (stuck rollout re-encodes on each retrieval); Psychosis = NA collapse of E1/E2 constraint + DA aberrant salience on noisy input. Extends MECH-027 (pathological modes) and MECH-080 (rollout truncation profiles) to PTSD/depression/psychosis and grounds all profiles in four-plane hierarchy. Generates intervention-order predictions (MECH-087): address upstream plane failures before # Registered 2026-03-15 — imagination / residue write gate cluster # Source: phenomenological analysis of imagination vs experience distinction - id: MECH-094 title: "Hypothesis tag is a categorical write gate separating simulation from committed residue updates; tag loss is the PTSD mechanism." claim_type: mechanism_hypothesis subject: default_mode.hypothesis_tag_residue_write_gate polarity: asserts status: candidate claim_level: functional evidence_quality_note: | EXQ-140 FAIL/weakens (2026-03-29): hypothesis tag gate discriminative pair FAIL. Consistent with prior finding that hypothesis tag gate as implemented is invalid as an approach. Weakening evidence accumulating. depends_on: - ARC-014 - INV-011 - INV-019 - MECH-060 - MECH-029 location: docs/architecture/default_mode.md#mech-094 source: - docs/architecture/default_mode.md - docs/architecture/control_plane_heartbeat.md notes: > The hypothesis tag — marking internally generated trajectories as non-committal — is the primary mechanism implementing ARC-014 constraint 3 (no residue update during Default Mode). The tag is categorical, not a precision attenuation parameter: it routes simulation signals away from the φ(z) write pathway entirely. Attenuation of φ(z) readout during imagination is a secondary effect; route separation is primary. Mechanism: during DMN/simulation, only the pre-commit error channel (MECH-060) is active; the post-commit realized-outcome channel — which drives residue accumulation — is suppressed. The qualitatively different affective texture of imagination vs. real experience is the phenomenological signature of this channel suppression: the absence of the full somatic/post-commit signal. Implements INV-011 (imagination without belief update) and INV-019 (rehearsal/durable-update separation) at the mechanism level. Pathological failure: when the hypothesis tag is lost (PTSD flashback, psychosis), the post-commit channel opens during what is structurally a replay event, and φ(z) begins accumulating residue from simulated/replayed content — the mechanism of MECH-076 (map structural deformation). The damage in rumination is tag loss, not excessive vividness. Simulation-driven viability map updates (MECH-092) are safe precisely because replay carries the hypothesis tag throughout. # Registered 2026-03-15 — control plane heartbeat architecture cluster # Source: conversation developing heartbeat model from MECH-057a FAIL analysis - id: ARC-023 title: "Three BG-like loops operate at characteristic thalamic heartbeat rates." claim_type: architectural_commitment subject: basal_ganglia.three_loop_thalamic_heartbeat polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | Held (2026-03-28): specific blocker is SD-006 (async multi-rate loop execution). SD-006 is not yet implemented beyond time-multiplexed phase 1. Full asynchronous multi-rate testing requires SD-006 phase 2 implementation. EXQ-131 FAIL/mixed (2026-03-29): multirate heartbeat discriminative pair FAIL. SD-006 phase 2 async required for full three-loop rate separation. Confirms existing hold. EXQ-131 reclassified INCONCLUSIVE (2026-03-30 review): var_harm_eval_on=1.05e-7 (near-zero) vs ablated=0.00276 -- E3 output freeze artifact of time-multiplexed (not async) substrate. Synchronous polling creates stale E3 state; does not replicate biological async multi-rate. Re-run (T215546Z) identical metrics -- one result. SD-006 phase 2 (true async execution) is a structural prerequisite. Does not weaken ARC-023. claim_level: mixed functional_restatement: > Three loops operate at distinct update rates: E1 continuous/frame-rate, E2 motor-command rate, E3 deliberation rate (slowest). An update-rate management mechanism prevents loop drift under variable processing latency. The functional requirement is rate-separated asynchronous updates regardless of substrate; the biological substrate is thalamic pacemaking, which is not required for the ANN implementation. depends_on: - ARC-021 - MECH-069 location: docs/architecture/control_plane_heartbeat.md#arc-023 source: - docs/architecture/control_plane_heartbeat.md notes: > Each BG-like cortico-striatal loop has a characteristic thalamic heartbeat rate: E1 at sensory frame rate (continuous); E2 at motor command rate; E3 at deliberation rate (slowest). Thalamic substrate: MD → E3 complex, VL → E2 motor pathway, sensory relay nuclei → E1. TRN acts as inter-loop routing switch. Coherence mechanism: thalamic clock prevents loop drift under indeterminate processing latency. Completion events are high-salience updates within the continuous stream, not the exclusive trigger. V1/V2 substrates use synchronous single-timestep updates and cannot test this claim (SD-006 required for V3). - id: MECH-089 title: "E1 updates are batched into theta-cycle summaries before reaching E3 (cross-frequency temporal packaging)." claim_type: mechanism_hypothesis subject: multirate.fast_to_slow_temporal_batching polarity: asserts status: active implementation_phase: v3 claim_level: mechanistic functional_restatement: > Fast loop outputs are temporally batched before slow loops sample them. E3 receives integrated summaries of recent E1 activity rather than raw frame-by-frame prediction errors. E3's minimum harm attribution resolution equals its own update window size — it cannot attribute finer-grained than one E3 update cycle. In an ANN substrate this is implemented as a rolling aggregation over E1 outputs before E3 sampling, not as oscillatory coupling. The biological mechanism (theta-gamma CFC) is one implementation of this functional requirement; it is not required in the ANN substrate. evidence_quality_note: | EXQ-052b PASS (2026-03-21, 5/5): ThetaBuffer confirmed filling with E1 outputs and batching them before E3 sampling. max_theta_buffer_size correctly recorded via _z_world_buffer attribute. E3 functional at ~1/9 the step rate (e3_tick_ratio ≈ 0.109), cal_gap=0.846. Demonstrates that temporal batching of fast-loop outputs before slow-loop sampling is structurally in place and metrically verifiable. Single PASS sufficient to elevate from provisional to active -- mechanism is implemented and measurable. EXQ-066 FAIL (2026-03-23, 4/5): BACKWARDS RESULT. e3_pred_error_batched=0.009 vs raw=0.004 -- batched E3 error 2.28x WORSE than raw. C1 fails. Interpretation: theta batching with theta_k=4 destroys temporal resolution E3 is using. The environment has meaningful state changes every step; a 4-step average loses that fine-grained signal. Two possible readings: (1) theta_k=4 is too coarse for this environment -- the right batch size may be 1-2 steps, consistent with the user's observation that theta rhythms need to be faster than tactical decision timescales; (2) context-gated batching (only during stable/uncommitted windows, not uniformly) may be needed. The backwards result is scientifically informative: E3 is using step-by-step temporal information, not benefiting from smoothing. Redesign should test context-dependent theta window sizes or shorter theta cycles (theta_k=1 or 2). EXQ-047g queued for SD-005 functional separation test; MECH-089 redesign deferred pending theta_k calibration. EXQ-122 FAIL (2026-03-28): theta_k=2 discriminative pair (ON vs ABLATED, 2 seeds). ADVERSE DIRECTION: harm_auc_ON=0.489 vs harm_auc_ABLATED=0.625 (delta=-0.135). Even with k=2, theta batching hurts harm attribution. C1/C2/C3 fail; C4/C5 pass (32 harm steps -- data quality OK). Strengthens context-gated interpretation: uniform temporal smoothing is harmful at any tested batch size. Genuine weakening for static-k variant. Hold at active; context-dependent gating design needed before promotion. depends_on: - ARC-023 - ARC-021 location: docs/architecture/control_plane_heartbeat.md#mech-089 source: - docs/architecture/control_plane_heartbeat.md notes: > Cross-frequency coupling (theta-gamma nesting) packages fast E1 updates for slow E3 consumption. Gamma cycles (E1-rate) nest within theta cycles (E3-rate); each theta cycle integrates ~5-7 gamma sub-cycles into a rolling summary that E3 samples at its heartbeat rate. E3 never receives raw E1 prediction error — only theta-cycle-averaged context. This defines the minimum temporal resolution for E3 harm attribution (cannot attribute finer-grained than one E3 heartbeat cycle). Thalamic relay performs the cross-rate integration. Analogous to theta-gamma coupling in hippocampal-prefrontal circuits. - id: MECH-090 title: "BG-level beta oscillations gate E3-to-action-selection propagation, not E3 internal updating." claim_type: mechanism_hypothesis subject: control_plane.commitment_gated_policy_output polarity: asserts status: active implementation_phase: v3 claim_level: mechanistic functional_restatement: > During committed action sequences, E3 continues updating its internal model but propagation of updated model state to action selection is gated. The gate opens at completion events or urgent interrupts. An urgent interrupt pathway can force an early gate-open without waiting for sequence completion. In an ANN substrate, this is implemented as a commitment-gated routing variable decoupling E3 forward pass from policy output — not as beta oscillation power. The biological mechanism (beta in STN/striatum) is one implementation; it is not required in the ANN substrate. evidence_quality_note: | EXQ-049 FAIL (2026-03-20): Same bug as EXQ-048 — agent.select_action() bypassed, gate never exercised. EXQ-049b fixes this. EXQ-059b FAIL (2026-03-20, 2/5 criteria): Same finding as MECH-057b — routing through select_action() restored but mean_running_variance=0.000. BetaGate elevation requires _running_variance to be populated, which only happens in the E3 training-loop path, not the inference path called during eval. C3 (beta elevated during committed sequence) and C4 (propagation blocked) untestable until variance accumulation is wired into the inference path. See MECH-057b for root-cause analysis. Fix: EXQ-049c calls agent.e3.update_after_execution(z_world_t, harm_prev<0) each step. EXQ-049b/c FAIL (2026-03-21): Chicken-and-egg deadlock — post_action_update() guards update_running_variance() behind _committed_trajectory is not None, but trajectory only set when committed, which requires variance < threshold. Variance frozen at 0.5 > 0.40. EXQ-049d FAIL (2026-03-21): Deadlock fixed by calling update_running_variance(wf_err) directly after world_forward loss. Committed condition PASS: hold_concordance=1.0, hold_count=10347. But over-correction: variance collapsed to 2.5e-6, agent permanently committed, n_uncommitted_steps=0 — C2 (uncommitted_release_concordance) has no data. Root cause: single-agent eval conflates training and eval conditions. Fix: two-condition design (EXQ-049e) — compare trained agent (low variance, committed, gate elevated) against untrained/reset agent (high variance, uncommitted, gate released). C1 (gate elevates when committed) is confirmed. C2 awaits EXQ-049e. EXQ-049e PASS (2026-03-21, 5/5): Two-condition design resolves the deadlock. Trained agent: 10,000 committed steps, hold_concordance=1.0, hold_count=10,347, variance=2.5e-6. Fresh agent: 321 uncommitted steps, release_concordance=1.0, propagation_count=50. Both conditions at stable extremes -- committed gate elevates, uncommitted gate releases. MECH-090 fully confirmed. EXQ-062b PASS (2026-03-22, 5/5): Surprise-gated spike selectivity -- confirms that committed gate can be re-opened by selective interrupt (surprise-gated variance spike) without requiring full de-commitment from all harm contacts. Corroborates the gating architecture -- gate holds during expected harm, opens only on genuine surprise. literature_evidence: | Jenkinson & Brown (2011) PMID 22018805 [DOI: 10.1016/j.tins.2011.09.003]: "Beta oscillations provide a measure of the likelihood that a new voluntary action will need to be actuated." Elevated beta = stay the course. Beta is predictive (modulated by dopamine in anticipation of whether action change is needed) not reactive. Loss of dopamine (Parkinson's) annuls prospective action resourcing. This is the clearest literature statement of MECH-090: beta gate signals status quo maintenance, release signals openness to policy update. Frank (2005) PMID 15701239 [DOI: 10.1162/0898929052880093]: Dynamic dopamine modulation in BG: D1 (direct/Go) and D2 (indirect/NoGo) are differentially modulated by phasic dopamine. Phasic dopamine during reward potentiates D1 (facilitates commitment); dopamine dip relieves D2 suppression (enables de-commitment). This is the mechanistic basis for MECH-106 hysteresis. Frank, Seeberger & O'Reilly (2004) PMID 15528409 [DOI: 10.1126/science.1102941]: "By carrot or by stick": Go/NoGo learning asymmetry dopamine-dependent. Confirms the direct/indirect pathway architecture and asymmetric commitment modulation. Redgrave, Prescott & Gurney (1999) PMID 10362291 [DOI: 10.1016/s0306-4522(98)00319-4]: BG as centralized selection mechanism: winner-takes-all via lateral inhibition in striatum resolves competition for access to motor and cognitive resources. Grounds the single-selection (not graded) nature of the commitment gate. depends_on: - ARC-023 - ARC-021 - MECH-057a location: docs/architecture/control_plane_heartbeat.md#mech-090 source: - docs/architecture/control_plane_heartbeat.md notes: > BG-level beta oscillations (~13-30 Hz, STN/striatum) gate propagation of E3 model updates to action selection -- NOT E3 internal model updating. During committed action sequence: beta elevated in BG -> E3 updates internally (heartbeat continues) but does not change action selection. At completion or hyperdirect stop-change signal: beta drops -> E3 model state propagates -> policy can change. Reframes MECH-057a: completion events are the principal policy-update opportunities within an otherwise beta-gated continuous E3 stream, not the exclusive E3 update trigger. Hyperdirect pathway (cortex->STN->GPi) provides fast interrupt for urgent stop-change without waiting for completion. Explains MECH-057a FAIL: experiment modelled binary gate (E3 blind until completion), not beta-gated propagation within continuous stream. SCOPE REFINEMENT (2026-04-08, from lit-pull): The output-only gating claim applies specifically to BG-level beta (STN/striatum). Cortical beta in premotor areas (SMA) also gates internal plan representation stability (Hosaka et al. 2016, Cerebral Cortex), meaning beta serves different gating functions at different circuit levels. Additionally, post-movement beta rebound carries internal model confidence information (Tan et al. 2016, J Neurosci), so the BG gate is not a dumb switch but an information-carrying signal. The claim's specificity -- output gating without internal gating -- holds at the BG level but does not generalise to all beta-generating circuits. - id: MECH-091 title: "Salient events phase-reset the E3 heartbeat clock." claim_type: mechanism_hypothesis subject: control_plane.salient_event_cycle_resync polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | Held (2026-03-28): specific blocker is SD-006 (async multi-rate loop execution). Phase-reset of E3 heartbeat requires SD-006 phase 2 async implementation first. EXQ-133 FAIL (2026-03-29): 5656 phase resets fired successfully but produced no discriminative behavioral signal (gap_on=0.014, criterion needs >=0.04). Resets execute but downstream async update consequences do not propagate. SD-006 phase 2 (fully async multi-rate loop) required for async effects to be observable. Investigation queued for why phase resets produce no downstream behavioral change even with current sync implementation. claim_level: mechanistic functional_restatement: > Salient events (completion, unexpected harm, commitment boundary) resynchronize E3's update window to start fresh rather than continuing mid-cycle. This ensures harm estimates from one sub-plan are fully integrated before the next planning window opens, preventing partial integration artefacts. In an ANN substrate, implemented as an explicit cycle-boundary marker triggered by salient events — not as oscillatory phase reset. The biological mechanism (thalamic-driven phase reset of theta/alpha oscillations, P300 substrate) is one implementation; it is not required in the ANN substrate. depends_on: - ARC-023 - MECH-090 location: docs/architecture/control_plane_heartbeat.md#mech-091 source: - docs/architecture/control_plane_heartbeat.md notes: > High-salience events (action completion, unexpected harm, commitment boundary crossing) phase-reset the E3 heartbeat oscillator — reorganise timing of subsequent cycles, not only boost signal amplitude. Phase reset ensures updated harm estimates from the salient event enter E3 at the start of a fresh cycle (no mid-cycle integration artefacts). In multi-step plans (W0→W1→W2) each waypoint completion phase-resets E3's clock, ensuring harm estimates from each sub-plan are fully integrated before the next begins. Neural basis: P300 event-related potential and its theta/alpha oscillatory substrate. - id: MECH-092 title: "Quiescent E3 heartbeat cycles trigger hippocampal SWR-equivalent replay for viability map consolidation." claim_type: mechanism_hypothesis subject: hippocampal.quiescent_offline_replay polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | Held (2026-03-28): specific blocker is SD-006 (async multi-rate loop execution). Quiescent heartbeat replay requires SD-006 phase 2 async loop before testing. EXQ-136 FAIL/weakens (2026-03-29): quiescent replay discriminative pair FAIL. SD-006 phase 2 async required for full offline consolidation. Confirms existing hold. claim_level: mixed functional_restatement: > During quiescent periods between active E3 updates, the hippocampal module performs offline replay of compressed recent trajectory experience to consolidate the viability map. This offline batch consolidation supplements online map updates and integrates experiences that occurred too fast for real-time E3 processing. This is standard experience replay — implementable in any substrate. The SWR mechanism (internally generated CA3 recurrent dynamics) is the biological substrate; the functional requirement is offline batch consolidation during idle periods, not the oscillatory mechanism. depends_on: - ARC-023 - ARC-014 - ARC-018 - ARC-022 - MECH-094 location: docs/architecture/control_plane_heartbeat.md#mech-092 source: - docs/architecture/control_plane_heartbeat.md notes: > During quiescent E3 heartbeat cycles (no new completion event or salient update), hippocampal module performs SWR-equivalent replay: replaying compressed recent trajectory experience (~10-20× faster than real-time) to consolidate viability map. Internally generated by CA3-equivalent recurrent dynamics, not requiring external input. V2 ARC-018 FAIL is partly explained by absence of any replay mechanism — map updates only at each discrete step, no quiescent consolidation phase. ARC-018 may be testable in V3 via explicitly modelled quiescent replay phases. Connects to sleep consolidation docs (docs/architecture/sleep/) as the within-session equivalent of sleep replay. Safe because replay carries the hypothesis tag (MECH-094) throughout: viability map consolidation updates the cognitive map of φ(z) without writing to φ(z) itself. This is the micro-quiescence analogue of ARC-014 (DMN): same replay machinery, operating at inter-action timescales rather than full task-disengagement timescales. - id: MECH-093 title: "z_beta modulates E3 heartbeat frequency, distinct from precision-weighting." claim_type: mechanism_hypothesis subject: affective.z_beta_e3_update_rate_modulation polarity: asserts status: candidate implementation_phase: v3 claim_level: functional depends_on: - ARC-023 - MECH-069 - MECH-059 location: docs/architecture/control_plane_heartbeat.md#mech-093 source: - docs/architecture/control_plane_heartbeat.md notes: > z_beta (affective latent) modulates E3 heartbeat frequency: high harm salience → faster E3 updates (finer temporal harm attribution); routine operation → slower E3 updates (more stable policy). Distinct from precision-weighting (MECH-059): precision-weighting scales how much each update matters; rate modulation scales how often updates occur. Corresponds to NA arousal axis in four-plane hierarchy (MECH-087): NA modulates perceptual sampling (MECH-084) and, via z_beta, E3 heartbeat rate. Predicted: sustained high harm salience → more reactive policy; routine conditions → increased policy stability. Scope clarification: MECH-093 governs E3 heartbeat frequency (arousal-modulated thalamic pacemaking) — it does NOT govern the respiratory plan-sweep clock (MECH-112). Heartbeat = arousal-level gating of E3 refresh rate; respiratory rhythm = periodic sweep of pre-commit hippocampal rollouts and DMN simulations for trajectory abandonment. These are distinct clocks at distinct timescales operating on distinct planning functions. evidence_quality_note: | EXQ-097 FAIL 2/3 (2026-03-26): C1 FAIL -- p1_rate_gap=-0.74 (threshold >= 2.0). E3 heartbeat rate did NOT differentiate high-harm from low-harm episodes; gap is negative (higher harm -> slightly lower rate) -- opposite of prediction. DIAGNOSTIC (2026-03-27): z_beta IS wired to clock rate. update_e3_rate_from_beta() is called every step in agent._e1_step() (agent.py:272). Implementation gap does NOT exist. Root causes: (1) z_beta does not encode harm salience sufficiently after 200 warmup episodes -- magnitude is not elevated during harm events. Pool skewed 16:1 (high-harm vs low-harm ticks), confirming harm is pervasive so z_beta has minimal between-episode variance. (2) Episode-level averaging is too coarse: the correct operationalization is step-level Pearson r(|z_beta|, 1/e3_steps_per_tick), not the mean clock rate gap between episode types. EXQ-097b redesigned with step-level metric + 500 warmup episodes. Hold candidate. EXQ-097b PASS 2/3 (2026-03-27): C1 r=0.964 (n=29937 steps) -- step-level z_beta/rate correlation confirmed (core claim). C3 action_var_ratio=0.096 -- 10x lower action variance in safe phases when BETA_MOD_ON. C2 harm_rate FAIL (ratio=1.58, higher in BETA_MOD_ON): state selection artifact -- high-z_beta phases are high-harm contexts by construction. Agent encounters more harm because it IS in high-harm situations, not because slowing causes harm. PASS = C1 AND C3. MECH-093 supported. Promote to provisional pending multi-seed replication. # Registered 2026-03-17 — SD-005 anatomical gap fill: TPJ, dual-stream encoder, peripersonal space # Source: neuroanatomical review of self/world latent split missing pieces - id: MECH-095 title: "Temporoparietal junction (TPJ) acts as agency-detection comparator distinguishing self-caused from other-caused change." claim_type: mechanism_hypothesis subject: tpj.agency_detection_comparator polarity: asserts status: active adjudicated_at_utc: "2026-03-25T00:00:00Z" adjudication_outcome: promoted implementation_phase: v3 evidence_quality_note: | EXQ-121 FAIL (2026-03-28): AUC_COMP_ON=0.411 vs AUC_COMP_ABLATED=0.745 (delta=-0.334). Only 1/5 criteria passed. AgencyComparator actively hurts attribution performance at current training scale. Possible causes: z_self (32-dim) too noisy for reliable efference-copy prediction; insufficient action embedding granularity; env_drift_prob=0.3 may be too low to produce enough discriminable other-caused events. Holds at active status -- prior neuroanatomical and literature support remains valid. Redesign flagged: larger z_self, improved efference-copy embedding, or higher env_drift_prob needed before next experimental test. depends_on: - SD-005 - MECH-069 - ARC-021 location: docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md source: - docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md notes: > At the z_self/z_world interface, an explicit agency-detection comparator (TPJ-equivalent) must compare efference-copy-predicted z_self change with observed z_self change. When predicted and observed z_self changes match, state change is self-caused (no residue). When they diverge, cause is attributed to z_world (potential residue). This is the architectural mechanism for SD-003 counterfactual attribution: self_delta and world_delta are only cleanly separable if this comparator exists. Without it, SD-003's causal_delta conflates proprioceptive self-effects with world-directed footprint. Neural basis: temporoparietal junction (TPJ) computes forward-model prediction vs sensory outcome mismatch for agency attribution. Blakemore et al. (2002). Additional grounding (2026-03-17): TPJ sits at the terminus of the lateral/third visual stream (MT→MST/FST→posterior STS→TPJ), which is specialised for dynamic motion and biological motion detection (Haak & Beckmann 2018; Pitcher & Ungerleider 2021). MSTd congruent/incongruent neurons (Gu et al. 2008) provide the upstream causal inference that feeds this stream: congruent=reafference, incongruent=exafference/world-change. The efference copy reaches parietal via SLF (Rolls et al. 2023). See MECH-098 and MECH-099. Literature: evidence/literature/targeted_review_reafference_streams/. evidence_quality_note: | EXQ-047k PASS 5/5 (2026-03-25, 4 seeds): contact_recall_routed=0.796 > 0.55, recall_improvement=+0.065 > 0.04, action_dissoc=-0.007 > -0.05, n_contact_min=164. 4-seed replication with clean discriminative test. Promoted from provisional to active. Combined with lit_conf=0.907 (Blakemore 2002, Gu 2008, Pitcher & Ungerleider 2021, Rolls 2023), evidence for TPJ as agency-detection comparator is strong. - id: MECH-096 title: "Dual-stream observation routing sends exteroceptive input to both E1 and a dedicated harm detection pathway." claim_type: mechanism_hypothesis subject: observation_encoder.dual_stream_routing polarity: asserts status: candidate evidence_quality_note: | V3 gate cleared (2026-03-28): SD-005 (z_self/z_world split) and SD-007 (reafference predictor) now implemented in ree-v3. Dual-stream routing experiments can proceed. EXQ-130 FAIL (2026-03-29): dual-stream routing pair FAIL. First experimental entry. EXQ-130 reclassified INCONCLUSIVE (2026-03-30 review): gap_stream_on=0.015, gap_ablated=0.019, delta=-0.004. Both gaps sub-threshold; 8x8 task too simple to discriminate dual-stream advantage (single-stream saturates task capacity). Small negative delta within noise at this scale. Re-run (T215504Z) identical metrics -- one result. Evidence direction corrected to inconclusive: experiment underpowered, not evidence against MECH-096. depends_on: - SD-005 - MECH-069 location: docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md source: - docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md notes: > The observation encoder must implement two anatomically motivated output heads, not a single encoder with a learned routing gate. Dorsal-equivalent head: egocentric, action-relevant, high temporal resolution -> z_self. Ventral-equivalent head: allocentric, object-identity, sustained representation → z_world. The two streams process structurally different information at different temporal resolutions and must remain architecturally distinct or the z_self/z_world split degrades back toward z_gamma conflation during end-to-end training. SD-005's "encoder routes to appropriate stream" phrase underspecifies this requirement. Neural basis: Goodale & Milner (1992) dorsal/ventral visual stream distinction. Update (2026-03-17): classical two-stream model is insufficient. HCP connectome data (Haak & Beckmann 2018, n=470) confirms a THIRD lateral stream (MT→MST/FST→posterior STS) specialised for dynamic motion and agency. This stream terminates near TPJ and is distinct from both dorsal and ventral at the white-matter-tract level. MECH-096 subject should be updated to three-stream encoder: dorsal head (→z_self), ventral head (→z_world content), lateral head (→agency/harm signal to E3). See MECH-099. Literature: evidence/literature/targeted_review_reafference_streams/. - id: MECH-097 title: "Peripersonal space geometry defines the commit locus — the spatial boundary of committed action." claim_type: mechanism_hypothesis subject: peripersonal_space.commit_locus polarity: asserts status: candidate evidence_quality_note: | V3 gate cleared (2026-03-28): SD-005 (z_self/z_world split) and ARC-016 (dynamic precision) now implemented in ree-v3. Peripersonal space / commit locus experiments can proceed. Note: depends on MECH-091 (salient event cycle-reset) which still needs SD-006 (async heartbeat); design experiments around the commit boundary itself independent of heartbeat reset if possible. EXQ-137 FAIL (2026-03-29): PPS commit locus discriminative pair FAIL. First experimental entry. EXQ-137 reclassified INCONCLUSIVE (2026-03-30 review): gap_locus_on=0.003, gap_ablated=0.008, delta=-0.005. SD-006 (async execution) and MECH-091 (salient event phase-reset) prerequisites unmet -- commit locus gate fires at wrong moments in synchronous substrate; phase-reset absorbed by stale E3. Experiment premature. Re-run (T215640Z) identical metrics -- one result. Does not weaken MECH-097. Requeue after SD-006 phase 2. depends_on: - SD-005 - MECH-091 - ARC-016 location: docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md source: - docs/thoughts/2026-03-14_self_world_latent_split_sd003_limitation.md notes: > The commit boundary coincides with the agent's peripersonal space (PPS) boundary: the dynamically-maintained near-body region where self-directed and world-directed action consequences first interact. Actions crossing this boundary (stepping onto a cell, contacting an object) constitute the class of z_self→z_world causal transitions that generate moral residue. In CausalGridWorld, contamination events occur exactly at this boundary: the agent's foot (z_self) makes contact with the cell (z_world). The PPS boundary is dynamically updated (scales with motor reach, attentional state) — this means the commit locus is not a fixed spatial threshold but a modulated one, consistent with z_beta affecting commitment thresholds (ARC-016). Without a PPS representation, the architecture has no spatial grounding for where self ends and world begins. Neural basis: Rizzolatti, Fogassi & Gallese (1997) premotor/parietal peripersonal space representation. VIP (ventral intraparietal area) confirmed as anatomical site via HCP tractography (Rolls et al. 2023, evidence/literature/targeted_review_reafference_streams/). downstream optimisation. # Registered 2026-03-17 — reafference cancellation and three-stream architecture - id: MECH-098 title: "Reafference cancellation: z_world subtracts predicted self-caused sensory change to isolate external events." claim_type: mechanism_hypothesis subject: encoder.reafference_cancellation polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-005 - MECH-095 - MECH-096 location: docs/thoughts/2026-03-17_three_stream_reafference.md source: - docs/thoughts/2026-03-17_three_stream_reafference.md notes: > The observation encoder must use its z_self prediction (efference copy from E2_self) to cancel expected perspective-shift components from z_world before E2_world operates. When the agent moves, the egocentric view shifts — this is a reafferent change fully predictable from the motor command. Only the residual after perspective-shift subtraction represents genuine world-state change. Without this cancellation, E2_world learns the identity shortcut (view shifted = world stable) rather than learning action-conditional world dynamics. The cancellation belongs at encoder level (before E2), not inside E2 itself. Biological basis: Area MSTd (Gu et al. 2008, Nature Neuroscience 11:1201–1210) contains congruent neurons (visual/vestibular aligned = reafference, world stable) and incongruent neurons (opposite = exafference, genuine world change). MSTd receives visual optic flow, vestibular, AND efference copy from premotor simultaneously — implementing exactly this decomposition at the early sensory processing stage. Cerebellar forward model (Wolpert & Kawato 1998) generates the predicted reafference; premotor→PPC pathway via SLF (Rolls et al. 2023) delivers efference copy to the parietal encoder stage. SD-003 consequence: after reafference cancellation, causal_sig = E3(z_world_residual_actual) - E3(z_world_residual_cf) will be sharp because only genuine world-content change contributes to the signal. The current SD-003 plateau at gap=0.027 (spurious BCE artifact) and collapse to 0.0007 (regression, EXQ-012) both result from the absence of this cancellation step. Literature: evidence/literature/targeted_review_reafference_streams/ (Gu 2008, Wolpert & Kawato 1998, Rolls 2023). EXQ-016 root-cause analysis (2026-03-18): V3-EXQ-016 tested a z-space ReafferencePredictor (SGD MLP, 200 steps) and achieved R2_test=0.118 vs the EXQ-014 lstsq benchmark of R2=0.333. Root causes: (1) encoder EMA alpha=0.3 means Δz_world is a blended multi-step signal — a single-step linear predictor from (z_self_t, a_t) cannot reconstruct it without the full EMA history; (2) C3 metric bug — the correction was applied to both endpoints of the delta, causing cancellation (dz_corrected = dz_raw to machine epsilon). SD-008 (fix alpha_world >= 0.9) is prerequisite for MECH-098 to be testable: with alpha_world=0.3 the reafference signal is too diluted for any single-step predictor. EXQ-021 retests with lstsq and fixed C3 metric. EXQ-023 retests with alpha_world=0.9. EXQ-022 (2026-03-20): lstsq reafference approach retired for this use case. R²_test=0.0 despite R²_train=0.373 -- extreme overfitting. Correction amplifies world-change signal 10x (dz_corrected=139.6 vs dz_raw=13.9). The lstsq model learns within-training-set correlations that do not transfer. Neural ReafferencePredictor (SD-007, validated EXQ-027b PASS) is the correct approach. All future reafference experiments use the neural predictor exclusively. EXQ-069 FAIL (2026-03-23): event_selectivity_delta=-0.002 -- reafference OFF is MARGINALLY more selective than ON. EXQ-082 FAIL: delta_approach_gap=-0.104 -- reafference ON is WORSE for approach detection. Root cause: predictor quality insufficient. reafference_r2_avg=0.339 (explains only 1/3 of z_world variance from locomotion). Residual correction error subtracts real signal along with noise. Current 2-layer MLP hidden_dim=64 architecture is the bottleneck. Upgrade path: 3-layer MLP with hidden_dim=128 or LSTM predictor. Both are feasible on available hardware (GTX 1050 Ti / RTX 2060 Super -- predictor is tiny relative to full model). R² target >= 0.70 before re-testing MECH-098. Note: the mechanism is architecturally correct (MSTd congruent neuron analog); the failure is engineering (predictor accuracy), not conceptual. Do NOT demote based on EXQ-069/082 alone. Redesign predictor first. EXQ-099 superseded (2026-03-27): both runs had collection filter bug (only 1 and 8 locomotion steps collected vs 2000+ expected; filter used ttype=='none' which misses hazard_approach/benefit_approach in CausalGridWorldV2). Superseded by EXQ-099a. EXQ-099a T223151Z PASS (2026-03-26): n_samples=50, gate_threshold=-1.0 (intentionally relaxed to test whether any correction level helps), R2_test=-0.249, selectivity_ratio= 1.655 >= 1.1, harm_avoidance_rate ON=OFF=1.0. Reafference correction produces cleaner selectivity even with a poor predictor. Mechanism confirmed conceptually. Predictor quality (R2 well below 0.70) remains engineering bottleneck for full validation. EXQ-099a T223306Z FAIL not_applicable (2026-03-26): n_samples=167, gate=0.70, R2=-0.027. Phase 1 gate failure -- engineering bottleneck persists at stricter quality threshold. - id: MECH-099 title: "Visual streams use a three-pathway architecture (ventral/dorsal/frontal) for object, action, and executive routing." claim_type: mechanism_hypothesis subject: visual_streams.three_pathway_architecture polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-017 - MECH-069 - MECH-096 location: docs/thoughts/2026-03-17_three_stream_reafference.md source: - docs/thoughts/2026-03-17_three_stream_reafference.md notes: > The visual processing hierarchy comprises three anatomically distinct pathways, not two (dorsal/ventral). Confirmed by Haak & Beckmann (2018, Cortex 98:73–83) from HCP resting-state data (n=470) with data-driven triple dissociation across 22 visual areas: (1) Dorsal (IPS/occipitoparietal): spatial processing, self-motion, visually guided action. Maps to z_self / SELF_SENSORY stream. (2) Ventral (VO/PHC): object identity, content. Maps to z_world / WORLD stream. (3) Lateral/third (MT→MST/FST→posterior STS): dynamic motion, biological motion, social/agency detection. Terminates near TPJ. Maps to the agency-detection channel (MECH-095 / E3 harm stream). These three pathways use anatomically distinct white matter tracts: SLF I/II/III for dorsal; ventral occipitotemporal white matter for ventral; a separate lateral pathway into STS for the third stream. This provides structural anatomical grounding for MECH-069 (incommensurability of error signals): the three learning channels are incommensurable partly because they run on physically distinct hardware. Collapsing them violates the biological architecture. MECH-096 (dual-stream encoder) must be updated to MECH-099-compliant three-stream encoder: dorsal head (→z_self), ventral head (→z_world content), lateral head (→agency/harm signal for E3). Literature: evidence/literature/targeted_review_reafference_streams/ (Haak & Beckmann 2018, Pitcher & Ungerleider 2021). evidence_quality_note: | EXQ-098 (2026-03-26): Two independent FAILs. Run 1 (seeds=42, 5 eps warmup): auc_delta=0.0000 (threshold >= 0.05). Run 2 (seeds=42+7, 400 eps warmup): auc_delta=-0.0384 -- lateral head UNDERPERFORMS baseline with more training. Adverse direction in run 2 suggests lateral head may actively interfere with the existing two-stream harm signal. Hold at candidate. EXQ-098b redesign needed before closing this question. The Haak & Beckmann three-pathway biological grounding is strong, but the current lateral head architecture (separate AUC scoring on harm channel) does not produce the expected signal improvement. EXQ-098b FAIL 2/3 (2026-03-27, 2-seed): Redesigned with agency attribution labels (agent_caused vs env_caused contact) + motion-delta input (MT analog). C2 PASS: attr_AUC_THREE=0.9942 >= 0.65 (both streams learn attribution well). C3 PASS: avoidance slightly higher with THREE_STREAM (+0.0007). C1 FAIL: auc_delta=-0.0047 (< 0.05 threshold) -- THREE_STREAM does not outperform TWO_STREAM. Key finding: z_world already encodes agency attribution nearly as well (AUC_TWO=0.9989) as the dedicated lateral head at this scale. The three-stream architectural advantage may only emerge at larger world_dim or with a deeper lateral head. Hold candidate. # Registered 2026-03-18 — EMA root-cause analysis and encoder supervision gap - id: MECH-100 title: "z_world encoder requires event-type cross-entropy auxiliary loss during training." claim_type: mechanism_hypothesis subject: encoder.event_contrastive_supervision polarity: asserts status: stable status_note: | Promoted provisional -> stable 2026-04-03. Evidence: conf=0.817, conflict_ratio=0, 4 exp all supports, 3 lit entries. EXQ-020 PASS (2026-03-20): selectivity_margin=0.882, event_classification_acc=0.692. EXQ-023 PASS (2026-03-22): combined SD-008+SD-009, event_selectivity_margin=0.084. EXQ-176/177 PASS confirming contrastive auxiliary loss is required for z_world discrimination. No conflicting experimental evidence. Stable mechanism; z_world encoder design is locked. implementation_phase: v3 depends_on: - SD-005 - SD-009 location: CLAUDE.md source: - evidence/experiments/ (V3-EXQ-013 event selectivity failure) notes: > The world encoder (z_world branch of SplitEncoder) requires an explicit event-type cross-entropy auxiliary loss during training. Standard reconstruction and E1-prediction losses are invariant to harm-relevance: they reward accurate decoding of all observation channels equally, giving the encoder no gradient signal to distinguish hazard-present from hazard-absent world states. As a result, z_world ends up representing general perspective-dominated content (90% variance explained by locomotion; EXQ-014 R2=0.333) with near-zero event selectivity (EXQ-013: Δz_world(env_hazard)≈Δz_world(empty), margin≈0.002). The fix (SD-009): add a small event classifier head (Linear(world_dim,3)) whose cross-entropy loss backpropagates through z_world during training. Labels: 0=none (empty locomotion), 1=env_caused_hazard, 2=agent_caused_hazard. This forces z_world to carry event-discriminative information, providing a stronger gradient for E2_world and E3 harm evaluation to exploit. Biological grounding: the ventral stream (MECH-099 stream 2) is trained by object-recognition supervision in development — not just reconstruction. Predictive coding alone does not produce object-discriminative ventral stream representations without categorical top-down signals (Murray et al. 2004, TICS 8:56–61). The event classifier is the minimal stand-in for such categorical supervision. Interaction with SD-008: the event contrastive loss will be ineffective if alpha_world=0.3, since the EMA smoothing dilutes the event contribution to 30% per step. Both SD-008 and SD-009 are co-prerequisites. Test: V3-EXQ-020 (event auxiliary loss alone), V3-EXQ-022 (combined with lstsq reafference), V3-EXQ-023 (SD-008 + SD-009 combined in alpha=0.9 run). EXQ-020 PASS on SD-009 mechanism (2026-03-20): selectivity_margin=0.882, event_classification_acc=0.692, Δz_world escalates env_hazard/agent_hazard 2.4× over none. C1 (calibration_gap) FAIL is inherited from SD-010 (broken z_harm evaluation), not from SD-009 itself. SD-009 mechanism validated — contrastive loss forces z_world to carry event-discriminative information. C1 retest pending SD-010 fix (EXQ-056c). - id: MECH-101 title: "z_world encoding requires reafference context to correctly attribute external vs self-caused change." claim_type: mechanism_hypothesis subject: encoder.reafference_world_context_required polarity: asserts status: provisional implementation_phase: v3 status_note: | Promoted candidate -> provisional 2026-03-22. Evidence: 3 supporting experiments, 0 weakening, conflict_ratio=0, conf=0.874. EXQ-027 (R2=0.027 with z_self input) established the failure mode. EXQ-021 PASS confirmed z_world_raw_prev input fixes the predictor. Implemented as SD-007 in ree_core/latent/stack.py. Literature triangulation pending (LIT-0090, LIT-0091) -- not required for provisional. depends_on: - MECH-098 - SD-007 location: docs/architecture/sd_004_sd_005_encoder_codesign.md#sd-007 source: - evidence/experiments/ # V3-EXQ-027 R²=0.027 failure analysis notes: > The SD-007 ReafferencePredictor must take z_world_raw_prev as input (not z_self_prev). In local-view environments, Δz_world_raw from locomotion includes cell content entering the field of view, which depends on current world state (available in z_world_raw_prev) but not on body proprioception (z_self). V3-EXQ-027 measured R²=0.027 with (z_self, a) inputs — near-zero because newly-revealed cell content dominates the delta and is inaccessible from body state alone. Biological basis: MSTd receives full visual optic flow (scene content-dependent) + vestibular + efference copy, NOT body state + efference copy. The optic flow encodes scene content; z_self encodes body position only. Fix: ReafferencePredictor(z_world_raw_prev, a) → Δz_world_loco. z_world_raw_prev is stored in LatentState and available in encode() as prev_state.z_world_raw. - id: MECH-103 title: "E1 performs multimodal exteroceptive fusion across sensory modalities before feeding E2." claim_type: mechanism_hypothesis subject: e1.multimodal_exteroception_fusion polarity: asserts status: candidate evidence_quality_note: | EXQ-128 FAIL (2026-03-29): superseded -- no multimodal sensory input in V3 substrate; artificial auditory channel not representative of genuine exteroceptive fusion. Manifest marked superseded. Claim untestable until V3 multimodal input is implemented. EXQ-134 FAIL (2026-03-29): somatosensory fusion pair -- same substrate issue; also superseded. Claim requires genuine multi-modality input streams. depends_on: - ARC-017 - ARC-004 location: docs/architecture/sensory_stream_tags.md#mech-103 source: - docs/thoughts/2026-03-19_sensory_stream_e3_decomposition.md notes: > Different exteroceptive modalities (vision, audition, somatosensory/touch) each contribute differently to E1's shared world latent (z_S / z_world). Multi-source convergence produces more accurate and robust world representations than single-modality input because each modality carries complementary structure: vision (object identity, spatial layout), audition (events, temporal cues, off-screen dynamics), somatosensory (surface properties, contact). Each modality has its own encoder pathway into the shared latent, contributing precision-weighted evidence. Grounds the WORLD stream tag's "modality fusion may occur upstream" in a concrete mechanism. Biological basis: superior temporal sulcus (STS) multisensory convergence; MST integrates optic flow with vestibular and auditory input; cross-modal ventral stream binding (Murray et al. 2004). - id: MECH-104 title: "Unexpected harm events spike commitment uncertainty (LC-NE volatility interrupt), enabling de-commitment." claim_type: mechanism_hypothesis subject: control_plane.volatility_interrupt polarity: asserts status: active implementation_phase: v3 status_note: | Promoted provisional -> active 2026-03-23. Evidence: conf=0.964, conflict_ratio=0, 3 supports (EXQ-049e, EXQ-064, EXQ-062b), 2 mixed (EXQ-062, EXQ-062a both criterion-revision iterations, not true weakening signals). Route-1 (jitter noise floor) confirmed (EXQ-049e PASS 5/5, EXQ-061 PASS). Route-2 (harm_surprise-gated spike) confirmed (EXQ-064 PASS 5/5): spike fires selectively on unexpected harm (delta_var_unexpected=0.051) and not on expected harm (delta_var_expected=0.000). EXQ-062b PASS 5/5 confirms spike selectivity (fewer spikes than always-spike condition, higher per-spike impact: 10.7 vs 6.7 uncommitted steps/spike). Both routes independently confirmed. Literature triangulation pending (LIT-0092, LIT-0093). Updated 2026-04-03: two additional definitive PASSes. EXQ-197 PASS 6/6 (2026-04-01): matched-seed discriminative pair, 3 seeds [42,7,13]. discriminative_delta=0.0506-0.0509 all seeds; ABLATED produces 0.0 (perfect discrimination). EXQ-204 PASS 4/4 (2026-04-03): surprise gate pair, 3 seeds. harm_spike_rv~1.6-1.9e-6 on unexpected harm; 0 on non-harm; 0 in ABLATED condition. Strongest evidence to date. claim_level: mechanistic functional_restatement: > When actual harm substantially exceeds predicted harm (prediction error surprise), the commitment uncertainty estimate (running_variance) receives an upward impulse proportional to surprise magnitude, enabling de-commitment from a currently active trajectory. Without this mechanism, world_forward training drives running_variance to near-zero, permanently locking the agent in committed state regardless of environmental change. In an ANN substrate this is implemented as a surprise-gated upward push to running_variance when |actual_harm - predicted_harm| exceeds a threshold. V3 approximation: Gaussian jitter (noise floor ε) added to the EMA so running_variance cannot fully collapse, allowing the committed→uncommitted transition to remain testable. The biological mechanism (LC phasic firing on unexpected outcomes, NE release increasing neural gain and driving exploration) is the reference implementation. See Yu & Dayan (2005) PMID 15944135. depends_on: - MECH-090 - ARC-016 - Q-007 location: docs/architecture/control_plane_heartbeat.md#mech-104 source: - docs/architecture/control_plane_heartbeat.md notes: > The locus coeruleus (LC) fires phasically on unexpected, behaviourally significant outcomes. NE release increases neural gain and drives exploration, reducing the effective commitment bias. This is Yu & Dayan (2005)'s unexpected uncertainty signal — distinct from expected uncertainty (routine variance tracked by the EMA), unexpected uncertainty is triggered by structural environmental change that the running average could not anticipate. Two V3 implementation routes: (1) jitter approximation — add small Gaussian noise floor ε to running_variance EMA update, preventing collapse to zero and making within-episode de-commitment testable; and (2) full wiring — harm_surprise = |harm_eval(z_harm) - actual_harm|; if harm_surprise > surprise_threshold, _running_variance += surprise_coeff × harm_surprise. Route (1) is sufficient for MECH-090/104 experiments. Route (2) is the proper mechanistic implementation for later V3 phases. Biological reference: Yu & Dayan (2005) Uncertainty, neuromodulation, and attention. Neuron 46(5):681-92. PMID 15944135. [DOI: 10.1016/j.neuron.2005.04.014] - id: MECH-105 title: "Hippocampal sequence completion drives BG beta release via subiculum-NAc-VTA dopaminergic loop." claim_type: mechanism_hypothesis subject: control_plane.hippocampal_bg_coupling_dopaminergic polarity: asserts status: candidate implementation_phase: v3 claim_level: mechanistic functional_restatement: > When the hippocampal module signals sequence completion (MECH-057b gate opens), this signal propagates to the BG beta gate (MECH-090) via a dopaminergic coupling loop: subiculum → nucleus accumbens → ventral pallidum → VTA → phasic dopamine release → D1 pathway potentiation → beta drop → gate release. This is the mechanistic bridge connecting the trajectory candidacy gate (hippocampal level) to the policy propagation gate (BG level). Without this coupling, the two gates are architecturally independent and sequence completion in HippocampalModule cannot trigger policy update in BetaGate. In an ANN substrate, the coupling is implemented as a completion signal from HippocampalModule to BetaGate mediated by a dopamine-analog variable modulating the commit threshold. The biological mechanism (subiculum→NAc→VP→VTA) is the reference. depends_on: - MECH-057b - MECH-090 - ARC-023 - ARC-021 location: docs/architecture/control_plane_heartbeat.md#mech-105 source: - docs/architecture/control_plane_heartbeat.md literature_evidence: | Lisman & Grace (2005) PMID 15924857 [DOI: 10.1016/j.neuron.2005.05.002]: Hippocampal-VTA loop: subiculum → NAc → ventral pallidum → VTA → dopamine release. Novel/completion events drive VTA via this pathway; upward arm is dopamine-enhanced hippocampal LTP. This is the precise anatomical substrate for hippocampal-to-BG completion signaling. Lansink et al. (2009) PMID 19688032 [DOI: 10.1371/journal.pbio.1000173]: Hippocampus leads ventral striatum in joint replay — hippocampal place cells fire before reward-correlated striatal cells. Directionality confirmed: completion signals originate in hippocampus and propagate forward to BG. Pennartz et al. (2011) PMID 21889806 [DOI: 10.1016/j.tins.2011.08.001]: Hippocampal-striatal axis: ventral striatum integrates hippocampal episodic content with value signals to generate goal-directed motivational predictions. Establishes functional as well as anatomical link. notes: > This claim resolves the architectural orphan problem: MECH-057b and MECH-090 are separately claimed but lack an explicit coupling mechanism. The subiculum→ NAc→VP→VTA loop is the missing link. In the V3 implementation, HippocampalModule must emit a completion signal received by BetaGate or E3Selector to trigger gate release. This coupling is currently absent and is a prerequisite for MECH-057b completion events to actually trigger MECH-090 gate release. Required changes: (1) HippocampalModule.generate_trajectories() emits completion signal on trajectory goal-reach or planned-length completion; (2) BetaGate or E3Selector receives this signal and calls gate.release(); (3) coupling is mediated by a dopamine-analog variable (see MECH-106). Without ARC-028 wiring, this claim cannot be tested experimentally. - id: MECH-106 title: "Commitment threshold is asymmetrically modulated by outcome valence: positive outcomes lower it, negative outcomes raise it." claim_type: mechanism_hypothesis subject: control_plane.commitment_hysteresis_outcome_valence polarity: asserts status: provisional implementation_phase: v3 evidence_quality_note: | EXQ-231 FAIL (2026-04-04): Superseded -- PERSISTENT and REACTIVE conditions produced identical commit values; experimental manipulation never created differentiated valence history. Design flaw, not claim failure. EXQ-231a PASS 5/5 criteria, 4/5 seeds (2026-04-04): VALENCE_BIAS vs NO_BIAS ablation. da_divergence=0.73-0.78 (4/5 seeds), threshold_asymmetry=2.7-2.8, latency_ratio=7-11x. Seed 42 anomaly (near-zero divergence, no commitments) is isolated; does not invalidate. Three literature supports (Frank 2005 D1/D2 asymmetry; Frank et al. 2004 Go/NoGo; Jenkinson & Brown 2011 beta modulation). conflict_ratio=0. Promoted to provisional 2026-04-06 governance. claim_level: mechanistic functional_restatement: > The threshold governing entry into and exit from committed states is asymmetric: positive outcomes (low harm, successful trajectory completion) lower the commit threshold via D1 pathway potentiation, making it easier to commit again. Negative outcomes (harm contact, large prediction error) raise the threshold via D2 pathway activation, making it harder to commit and easier to break from an active commitment. This asymmetric hysteresis prevents pathological flip-flopping (committed state is sticky through minor perturbations) while ensuring genuine harm events can force de-commitment. In an ANN substrate, the hysteresis is a slowly-updating outcome-valence bias on the commit threshold: commit_threshold_effective = commit_threshold_base × (1 + valence_bias), where valence_bias increases after harm and decreases after successful completion. The biological mechanism (phasic dopamine modulating D1/D2 balance in striatum) is the reference. depends_on: - MECH-090 - ARC-016 - MECH-104 - MECH-105 location: docs/architecture/control_plane_heartbeat.md#mech-106 source: - docs/architecture/control_plane_heartbeat.md literature_evidence: | Frank (2005) PMID 15701239 [DOI: 10.1162/0898929052880093]: Dynamic dopamine modulation in BG: D1 (direct/Go) potentiated by positive phasic dopamine — lowers threshold for action facilitation (commit). D2 (indirect/NoGo) relieved by dopamine dip — raises suppression signal (de-commit). Asymmetric dynamic range explains over-commitment (Parkinson's off medication) and under-commitment (dopamine overdose in intact striatum). Frank, Seeberger & O'Reilly (2004) PMID 15528409 [DOI: 10.1126/science.1102941]: "By carrot or by stick": Go/NoGo learning asymmetry is dopamine-dependent. Parkinson's patients off medication learn better from negative outcomes (D2 dominant); on medication the reverse (D1 dominant). Confirms asymmetric modulation of commitment direction by outcome valence. Jenkinson & Brown (2011) PMID 22018805 [DOI: 10.1016/j.tins.2011.09.003]: Beta oscillations are predictive — modulated by net dopamine levels at cortical BG inputs in response to salient cues. Loss of dopamine (Parkinson's) annuls the prospective resourcing of actions. Connects outcome-valence dopaminergic modulation to the beta gate mechanism (MECH-090). notes: > Commitment hysteresis is the BG-level mechanism preventing pathological state-switching. Without it, any variance fluctuation across the threshold produces an immediate state change. The dopaminergic asymmetry means successful sequences build commitment momentum (threshold lowers, harder to de-commit prematurely) while harm events build release momentum (threshold raises, easier to de-commit). This is slower-timescale than running_variance — it is a motivational bias rather than a prediction-error signal. Distinct from MECH-104 (volatility interrupt, fast) and running_variance (medium), this is the slow dopaminergic modulation that sets the operating point. Relevant to MECH-025 (action doing mode) because that mode requires low threshold + committed state, which the dopamine system builds through accumulated successful sequences. V3 implementation: valence_bias updated at each E3 tick via a slow EMA on outcome valence signal (harm_t vs predicted_harm_t). - id: ARC-028 title: "HippocampalModule must emit a trajectory completion signal wired to BetaGate to couple the candidacy and propagation gates." claim_type: architectural_commitment subject: control_plane.hippocampal_betagate_coupling_wiring polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-057b - MECH-090 - MECH-105 - ARC-023 location: docs/architecture/control_plane_heartbeat.md#arc-028 source: - docs/architecture/control_plane_heartbeat.md notes: > MECH-057b (hippocampal candidacy gate) and MECH-090 (BG beta propagation gate) are currently architecturally independent in V3. ARC-028 commits to coupling them: HippocampalModule must emit an explicit trajectory completion signal received by BetaGate (or the commit_threshold mechanism) to trigger gate release. Without this wiring, sequence completion in HippocampalModule has no downstream effect on BetaGate and the two gates cannot be tested as a unified circuit. Required implementation changes: (1) HippocampalModule.generate_trajectories() emits a completion signal when a trajectory reaches its goal or planned length; (2) E3Selector or BetaGate receives this and calls beta_gate.release(); (3) the coupling is mediated by a dopamine-analog variable (see MECH-105, MECH-106) for proper hysteresis. Biological basis: subiculum → NAc → VP → VTA → dopamine → BG beta release (Lisman & Grace 2005, PMID 15924857). - id: ARC-029 title: "Committed and uncommitted operating modes produce measurably distinct harm outcomes." claim_type: architectural_commitment subject: cognitive_modes.behavioral_consequence_of_commitment polarity: asserts status: provisional implementation_phase: v3 evidence_quality_note: | EXQ-125 FAIL (2026-03-29): committed mode discriminative pair FAIL. Adding weakening evidence. Prior EXQ-063 PASS remains valid. Mixed evidence building. EXQ-125 re-run FAIL 5/6 (2026-04-03): C1 FAIL -- harm_gap_stable = -0.01222 (threshold >= 0.0001). Committed mode WORSE than ablated in stable env (committed harm_rate higher). C2-C6 all PASS (gap_ratio=1.064, committed steps present, zero committed steps in ablated, no fatal errors). This is the opposite pattern from EXQ-063 (harm_gap_stable=+0.00041). Mixed evidence now: one PASS (EXQ-063) and two FAILs (EXQ-125 x2 runs). The committed mode may produce idiosyncratic harm differences that depend on random seed and training state rather than a systematic architectural advantage. ARC-029's core prediction (committed < ablated in stable) is not consistently confirmed. Hold at provisional; more seeds needed. EXQ-125 FAIL 3/6 (2026-04-04, seeds=[42,123]): seed 42 shows committed mode WORSE in stable (harm_gap_stable=-0.000091 < threshold); seed 123 shows it better (+0.000234). avg_harm_gap_stable=0.000071 (barely above zero, avg_gap_ratio=1.943). 3rd FAIL across EXQ-125 runs vs 1 PASS (EXQ-063). Diagnosis flagged: systematic difference between EXQ-063 and EXQ-125 designs unknown (seed set, warmup length, env config). /diagnose-errors requested before any status change. Maintain provisional pending diagnosis. EXQ-227 DIAGNOSTIC non_contributory (2026-04-05, cowork-2026-04-05-a): attempted to reproduce EXQ-063 exact conditions (seeds [0,1]). SET A harm_gap_stable=-3.26e-05 (gap sign inverted, fails bare>0 criterion). EXQ-063 PASS does not reproduce under current substrate. Root cause: SD-010/011/012 substrate changes reduced harm rates ~100x, making any committed-gate signal undetectable at original env parameters (num_hazards=4, hazard_harm=0.02). This confirms the EXQ-063 vs EXQ-125 discrepancy is substrate drift, not seed-dependence. EXQ-125a redesign (num_hazards=8, hazard_harm=0.05, 5 seeds, harm_obs passed to sense()) is the correct follow-up. /diagnose-errors complete. status_note: | Promoted candidate -> provisional 2026-03-23. Evidence: conf=0.774, 1 support (EXQ-063 PASS 5/5), 0 weakens, conflict_ratio=0. EXQ-063 is the gate-ablation 2x2 design described in notes (EXP-0085). harm_gap_stable=0.00041 (gate-active outperforms ablated in stable env), harm_gap_volatile=0.00023 (advantage narrows, ratio=0.559 -- commitment is context-dependent, not universally beneficial). Ablation confirmed complete. Both predicted effects confirmed. Single experiment; more evidence desirable before active promotion. depends_on: - ARC-016 - MECH-090 location: docs/architecture/modes_of_cognition.md#arc-029 source: - docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md notes: > The gate-ablation design (EXP-0085) is the cleanest test: train a single agent normally, then compare (A) normal BetaGate enabled vs (B) BetaGate ablated (always-release). Controls for training quality; isolates only the commitment gate's effect on behavior. Two environment conditions add a second dimension: stable hazard layout (commitment should help: agent holds a safe trajectory) vs volatile layout (hazard moves every N steps: commitment should hurt or be neutral as agent cannot adapt mid-commitment). PASS pattern: harm_committed < harm_ablated in stable env; gap narrows or reverses in volatile env. This would confirm that commitment is not universally good but context-dependent -- the correct functional prediction of ARC-016's control-plane framing. ENVIRONMENT REGIME NOTE (2026-04-05, from Humphries 2012 lit-pull): dopamine shifts BG output between exploitation (stable env) and exploration (volatile env). EXQ-125a uses num_hazards=8 -- this may be in the volatile regime where exploration (uncommitted mode) is dopaminergically optimal. If so, committed mode should show no advantage or active disadvantage in an 8-hazard environment, which would constitute a FAIL not of the claim but of the experimental regime. Consideration for follow-up: explicitly test the stable/volatile axis by varying hazard density (e.g. 2-3 hazards = stable/exploitative, 6-8 = volatile/exploratory) and predict that committed mode advantage is strongest at low hazard density. The Humphries framework predicts the interaction: commitment_advantage = f(stability), so the correct test is the stability x commitment interaction, not just the main effect. EXQ-125a tests only the high-hazard condition. - id: ARC-030 title: "The three BG-like loops require symmetric Go (approach) and NoGo (avoidance) sub-channels; pure NoGo architecture produces behavioral flatness." claim_type: architecture_hypothesis subject: architecture.approach_avoidance_symmetry polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | Hold at candidate (2026-03-29): no new experiments this session. SD-010 dependency and approach-avoidance symmetry (Go sub-channel implementation) required before testing. No experimental evidence yet. EXQ-086 INCONCLUSIVE/bug (2026-03-30, re-run of 20260323): benefit_rate=0 despite 274 buffered benefit events -- measurement bug in benefit_rate computation from buffer. Go sub-channel logging works (events recorded) but rate metric broken. Does not test ARC-030. Superseded by EXQ-086a (corrected benefit_rate computation, to be designed). EXQ-085m FAIL 0/4 (2026-03-30): benefit_eval_head in E3 (benefit proximity regression head trained jointly with E3). benefit_eval_r2_train=-8.6 (head diverges completely). benefit_per_ep=0 in both ENABLED/DISABLED conditions -- benefit events not occurring in 1-seed short warmup. ARC-030 approach-via-benefit-head not testable at this scale. Structural issue: benefit signal too sparse during warmup; head cannot learn proximity regression without sustained benefit contact. ARC-030 requires a richer benefit signal or pre-trained benefit head before E3 integration is testable. Goal-lift battery (2026-04-01): EXQ-183 FAIL 2/4 (BCE shared selector): benefit_eval_auc =0.538 (barely above chance); combined/random=0.73x. benefit_eval_head learns but does not discriminate effectively. EXQ-186 FAIL 1/4 (hybrid benefit+harm): C2 PASS (combined/ harm_only=1.55x -- adding benefit DOES improve over harm-only), but C1 FAIL (combined/ random=0.25x), C3 FAIL (harm 1.4x worse), C4 FAIL (combined < benefit_only -- harm avoidance is counterproductive). The C2 PASS is a positive signal for ARC-030: the shared selector does produce approach-avoidance integration lift vs pure avoidance. But the absolute performance is far below random baseline, suggesting 1-step greedy action selection on 10x10 grid is fundamentally insufficient. Oracle ceiling test (EXQ-182a) is the gate. EXQ-138a FAIL 2/5 x2 runs (2026-04-03): Go/NoGo symmetry discriminative pair FAIL. Run 1 (T04:36): criteria_met=2/5. Run 2 (T09:04): criteria_met=2/5. delta_benefit_rate=+0.00002 (threshold needs >=0.002). benefit_rate_go~=0.00039 (needs >= 0.002). C3 PASS (harm preserved). C5 PASS (benefit_buf_go >= 100 confirmed). The Go channel is wired and benefit events are recorded (n_benefit_buf_go ~750/seed) but rate is ~100x below threshold. Consistent result across both runs: Go sub-channel does not produce measurable behavioral lift in benefit acquisition at V3 scale with current training budget (warmup=400 eps). Weakening evidence for ARC-030 at this scale. EXQ-226 PASS 2/2 criteria (2026-04-04): FIRST PASS for ARC-030. COMBINED (harm+goal CEM) vs HARM_ONLY: benefit_ratio=1.319 (>= 1.2 threshold), harm_ratio=0.762 (<= 1.3 ceiling), avg_goal_norm=0.360 (precondition met). Key caveat: aggregate PASS driven by seed 13 where HARM_ONLY collapsed (0 resources, 29.5% harm); COMBINED rescued it (0.12 resources, 14.2% harm). Seeds 42 and 7 show COMBINED roughly equivalent to HARM_ONLY. The result supports ARC-030's structural prediction -- behavioral collapse under pure NoGo is a stochastic attractor, and the Go channel prevents it -- but the effect is intermittent, not universal. Critical design finding: fixed-weight CEM scoring (not a learned benefit_eval head) was what finally worked. All 5+ prior failures were head-learning failures, not architecture failures. V4 implementation of approach-avoidance should start from CEM scoring. GOVERNANCE META (2026-04-06): Illusory conflict resolution risk. Go/NoGo symmetry untested with real z_goal. Approach sub-channel cannot be evaluated when goal signal is absent. Non-contributory FAILs (EXQ-235, EXQ-086 measurement bug) removed from scoring but the remaining supports are from specialized selectors, not the full symmetric gate. Status: pending_retest_after_substrate (gate: EXQ-247 + SD-015 resource indicator encoding). depends_on: - ARC-021 - MECH-069 - INV-032 - SD-010 location: docs/architecture/approach_avoidance_symmetry.md#arc-030 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > The biological BG model motivating ARC-021 has explicit D1 (Go/approach) and D2 (NoGo/avoidance) pathways in competition within each loop. REE's current formulation specifies NoGo in architectural detail (nociceptive stream SD-010, agency comparator MECH-095, precision-to-commitment wiring ARC-016) but does not build out Go. Without approach drives, the gradient minimum under harm-avoidance-only training is quiescence: the agent that does nothing accrues no harm signal, no attribution, never crosses a commit boundary, and appears by its own error metrics to be performing optimally. The residue field requires positive attractors alongside harm repellers or the hippocampal planner has no terrain to navigate toward. The appetitive stream (goal-approach MECH-112, novelty MECH-111, self-maintenance MECH-113) must be structurally co-designed with the nociceptive stream (SD-010). Registered 2026-03-22. Literature: Cox et al. 2015 (NeuroImage, DOI 10.1016/j.neuroimage.2014.12.070) -- human PET confirms D1 binding predicts approach learning, D2 binding predicts avoidance learning; "bidirectional modulatory role for striatal dopamine via segregated D1 and D2 pathways." Hikida et al. 2012 (PNAS, DOI 10.1073/pnas.1220358110) -- D1 activation (direct pathway) = reward learning; D2 inactivation (indirect pathway) = aversive learning; pharmacological dissociation confirmed. Bariselli et al. 2018 (Brain Res, DOI 10.1016/j.brainres.2018.10.009) -- CRITICAL: COMPETITIVE MODEL: dMSNs and iMSNs are tuned to the SAME actions and compete to determine behavioral response. NOT two separate action inventories. Both pathways evaluate the same trajectory proposals from hippocampus; the commit threshold (MECH-106, ARC-016) is the competition balance point. Architecture shape: ARC-030 implementation is NOT two parallel pipelines selecting different trajectories. It is dual evaluation of the SAME hippocampal trajectory proposals: harm channel (SD-010/NoGo) and goal channel (MECH-112/Go) evaluate each proposal; the precision-weighted balance determines commitment. MECH-112 and SD-010 are symmetric inputs to the same E3 commitment gate. This extends MECH-106 (commitment threshold modulated by valence) and closes the ARC-016 precision-to-behavior wiring gap. Full synthesis: evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md - id: ARC-031 title: "HippocampalModule navigates z_self trajectory space (deliberation sequences) in addition to z_world action-object space O, enabling the agent to plan its own deliberation." claim_type: architecture_hypothesis subject: hippocampal.z_self_trajectory_navigation polarity: asserts status: candidate implementation_phase: v4 depends_on: - SD-004 - MECH-113 - MECH-114 - MECH-092 - ARC-028 - INV-033 location: docs/architecture/approach_avoidance_symmetry.md#arc-031 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md notes: > Current HippocampalModule (SD-004 / Q-020 resolved) navigates action-object space O in the z_world domain -- it maps and plans through the external world. ARC-031 extends this to z_self trajectory space: planned sequences of internal self-state transitions constituting deliberation, attention-shifting, or rumination. A z_self trajectory encodes "I will first resolve uncertainty in domain X (shift from high to low D_eff), then attend to Y, then commit." E3 evaluates these trajectories by self-coherence cost, hypothesis-tag integrity cost, and self-maintenance cost -- not harm cost. Biological basis: cognitive map literature shows hippocampal place/grid cells represent not just spatial location but position in any continuous abstract space. Behrens et al. 2018 (Neuron, DOI 10.1016/j.neuron.2018.09.040): grid/place cell algorithms map social cognition and conceptual structure -- navigation algorithms apply to abstract "thought space." Schapiro et al. 2017 (J Cogn Neurosci, DOI 10.1162/jocn_a_01111): hippocampal representations encode temporal sequences in abstract domains. Connection to ARC-028 (hippocampal->BG completion coupling): if hippocampus generates both external and deliberation trajectory proposals, the BG completion signal fires on BOTH action completion and deliberation completion. ARC-031 extends the completion coupling to include internal cognitive closure events. MECH-113 D_eff monitoring is the navigation PREREQUISITE: hippocampal self-navigation is only meaningful if z_self is coherent enough to define a stable navigable space. High D_eff = dispersed self-model = no reliable position in z_self space = navigation degenerates. MECH-114 D_eff commit gate must be in place before ARC-031 is testable. MECH-092 offline replay extended to z_self trajectories: during quiescent E3 cycles, consolidation of deliberation sequences (not just world-action sequences) occurs. V4 scope: requires waking architecture (EXQ-072..075) validated before implementation. See also sleep mechanisms (SWR consolidation of place-reward associations -- the self-trajectory analog would be deliberation-outcome associations consolidated in sleep). Registered 2026-03-23. - id: ARC-027 title: "Harm stream runs as a parallel thalamic pathway independent of world latent space." claim_type: architectural_commitment subject: harm_stream.parallel_thalamic_pathway polarity: asserts status: active implementation_phase: v3 depends_on: - ARC-017 - ARC-003 location: docs/architecture/sensory_stream_tags.md#arc-027 source: - docs/architecture/sensory_stream_tags.md evidence_quality_note: | EXQ-027b PASS (2026-03-19): Reafference diagnostic confirms SD-007 hurts E3 calibration when applied to z_world (correction_delta=-0.045, harm_pred_std drops from 0.108 to 0.008). Over-correction explained by applying reafference to a fused stream that includes the nociceptive signal. If z_harm were separate, reafference correction to z_world would not strip the harm signal. EXQ-044 FAIL (2026-03-19): SD-003 attribution collapsed again (world_forward_r2=0.0, all causal_sig ≈ 2.7e-5, E3 harm_eval=0.5 uniform). Sequential training fix was insufficient because the comparison is being made in z_world, which conflates nociceptive and exteroceptive content. Attribution needs to operate on the harm stream specifically, not the mixed world latent. EXQ-045 FAIL (2026-03-19): MECH-102 advantage ladder reversed (none=0.067 > approach=0.049 > contact=0.044). Ethical advantage is highest in safe states because action-space divergence is largest before hazard contact. Consistent with the harm-stream architecture: maximum ethical leverage is in the approach gradient, not at contact — which requires the harm stream to be available during trajectory planning, not only at contact events. EXQ-047 FAIL (2026-03-19): SD-005 split latent improves attribution by 82% (0.034 vs 0.019) but calibration gain (21%) falls below 5pp threshold. Split along z_self/z_world axis is insufficient — nociception needs to be a third separate stream, not subsumed in z_world. All four experiments converge: the current architecture fuses nociceptive content into z_world and this conflation is the root cause of SD-003/SD-007/MECH-102 failures. ARC-027 implementation is prerequisite for SD-003 to work. EXQ-058b ARC-027 PASS (2026-03-20, 5/5 criteria): Foundational V3 validation. HarmEncoder (2-layer MLP, harm_obs → z_harm, trained with direct MSE supervision on hazard/resource proximity labels) achieves: mean_harm_none=0.014, mean_harm_approach=0.726, mean_harm_contact=0.987; Pearson_r=0.973 (threshold 0.85); calibration_gap=0.712 (threshold 0.40); harm_pred_std=0.417 (not collapsed; threshold 0.05). Separate pathway confirmed: z_harm is learned from harm_obs alone, not from z_world; the encoder generalises across event types with strong rank ordering. This is the first clean demonstration that the harm stream architecture works when architecturally isolated. Prerequisite validation for SD-003 + SD-010 joint tests. EXQ-056c SD-010 PASS (2026-03-21, 4/5): HarmEncoder retest with reafference isolation confirmed. Pearson r_z_harm=0.914 (threshold 0.85). Harm stream produces clean proximity signal distinct from z_world, not subject to reafference correction. Confirms architectural separation is robust across experimental variations. notes: > The HARM stream must be implemented as a parallel sensory pathway with its own sensory input and training signal, not derived from z_world or E1 processing. It receives direct thalamic input (fast, coarse threat detection — the "low road") feeding the amygdala-equivalent harm-detection system before cortical world-model processing occurs. Consequences: (1) harm_eval is an input to E3, not a head inside E3; (2) SD-007 reafference correction applies at the HARM stream level, not at z_world — EXQ-027b over-correction is explained by applying reafference to the wrong signal; (3) SD-003 attribution compares HARM stream against E2 counterfactual rollout, not internally within E3; (4) EXQ-043/044 calibration collapses are explained: training a sensory detection system through a planning pipeline is architecturally incorrect. Biological basis: spinothalamic tract provides fast nociceptive input direct to amygdala (VPM/Po → amygdala) in parallel with slower cortical route (thalamus → S1 → insula → amygdala). LeDoux (1996) "low road" — amygdala rapid threat response depends on direct pathway, anatomically distinct from cortical world-model processing. Note: previously misregistered as ARC-025 (ID collision with architecture.three_engine_irreducibility). Corrected to ARC-027 (2026-03-19). # ── Design Decisions (SD) ──────────────────────────────────────────────────── - id: SD-001 title: "CEM-based trajectory search belongs in HippocampalModule, not E2." claim_type: design_decision subject: hippocampal_module.cem_trajectory_search_placement polarity: asserts status: resolved status_note: > CEM-based trajectory search was misplaced in E2. Resolved V2: CEM moved to HippocampalModule. E2 is now a pure transition model f(z_t, a_t) → z_{t+1}. depends_on: [] location: docs/architecture/hippocampal_systems.md - id: SD-002 title: "E1 associative prior must be wired into HippocampalModule for terrain-informed rollout search." claim_type: design_decision subject: hippocampal_module.e1_associative_prior_wiring polarity: asserts status: resolved status_note: > E1 associative prior must be wired into HippocampalModule terrain search (E1→HippocampalModule mutual constitution). Resolved 2026-03-06. depends_on: [SD-001] location: docs/architecture/hippocampal_systems.md - id: SD-003 title: "Self-attribution via counterfactual E2: causal signature = E2(z_t, a_actual) minus E2(z_t, a_counterfactual)." claim_type: design_decision subject: self_attribution.counterfactual_e2_pipeline polarity: asserts status: validated status_note: > EXQ-030b PASS (2026-03-18): validated on z_world pipeline -- causal_sig = E3(z_world_actual) - E3(z_world_cf), world_forward_r2=0.947, attribution_gap=0.035, correct sign structure. This validation predates SD-010 full wiring (E3 now takes z_harm, not z_world). The E2-counterfactual pipeline architecture is confirmed sound but the target stream has changed. Post-SD-011, the correct pipeline is: z_harm_s_cf = E2_harm_s(z_harm_s, a_cf); causal_sig = E3(z_harm_s_actual) - E3(z_harm_s_cf). EXQ-093/094 confirmed that HarmBridge(z_world -> z_harm) is infeasible (bridge_r2=0 architectural, SD-010 makes z_world perp z_harm). ARC-033 + SD-011 implementation required before SD-003 can be retested. evidence_quality_note: > Literature pull 2026-03-28 (4 entries, lit_conf=0.859, all supports) identified two caveats that must be resolved before V3 retest. (1) Biological precedent gap: Frith 2000 and Shergill 2003 directly evidence the single-pass comparator (E2(a_actual) - observed), not the two-pass counterfactual SD-003 implements. The two-pass extension is an inferential step with no direct biological evidence. (2) Interventional training requirement: Scholkopf 2021 establishes that a regression-trained forward model learns the correlational distribution P(z_harm_s | z_t, a), not the interventional distribution P(z_harm_s | do(a)). In confounded states (where agent presence correlates with ambient harm), E2_harm_s trained on observational trajectories alone will produce a biased causal_sig. ARC-033 (E2_harm_s) implementation must include counterfactual perturbation signal in training -- not only observational rollouts -- for SD-003 to be valid. EXQ-115 tests the new z_harm_s pipeline; a follow-up varying interventional vs observational training is needed if EXQ-115 passes. EXQ-166a FAIL 4/6 (2026-03-30): obs-space forward model (Approach B, 4 seeds, 80-ep P0). obs_fwd_r2 bimodal: seeds 42/seed3=0.925/0.885 (strong convergence), seeds 1/2=0.0 (P0 encoder quality failure despite extended warmup). C2 delta_approach=0.0006 (threshold 0.03): near-zero causal gap even for converging seeds. Architecture diagnostic: obs-space forward model learns general obs dynamics (position-change dominated) without selective pressure on harm-causal structure. Loss function indifferent to harm-relevant obs dims (diluted in 51-dim obs by position/resource dims). Proposed fix (EXQ-166b): z_harm_s latent forward model with reconstruction branch (autoencoder) to prevent identity collapse -- keeps learning focused on harm-specific structure. C2 failure is architectural, not a training budget issue. EXQ-195 FAIL (2026-04-04): attribution_gap=-0.044 all 4 seeds (C1 FAIL). Classified non-contributory (2026-04-06 governance): negative gap reflects missing SD-003 full pipeline wiring (counterfactual comparison not correctly structured), not claim failure. harm_forward_r2=0.914 (ARC-033 component working). Substrate gap, not claim refutation. depends_on: [SD-005, SD-007, SD-008, MECH-071, SD-011, ARC-033, SD-013] location: docs/architecture/sd_003_experiment_design.md - id: SD-004 title: "Action objects as hippocampal map backbone, enabling long-horizon planning beyond step-level trajectories." claim_type: design_decision subject: hippocampal_module.action_object_map_backbone polarity: asserts status: implemented status_note: > E2 produces compressed action-object representations; HippocampalModule navigates action-object space O rather than raw z_world. Enables planning horizons far beyond raw state space. SD-004 and SD-005 must be co-designed: action objects require z_world to exist. depends_on: [SD-005] location: docs/architecture/sd_004_sd_005_encoder_codesign.md - id: SD-005 title: "z_gamma split into z_self (E2 motor-sensory domain) and z_world (E3/Hippocampus domain)." claim_type: design_decision subject: latent_stack.self_world_latent_split polarity: asserts status: implemented status_note: > z_gamma split into z_self (E2 domain: motor-sensory, proprioceptive) and z_world (E3/Hippocampus/ResidueField domain: causal footprint, residue, moral attribution). SD-005 and SD-004 must be co-designed. evidence_quality_note: | EXQ-047 (2026-03-20): Split provides 82% attribution improvement (gap: 0.035 split vs 0.019 unified) and meaningful calibration gain (0.1915 vs 0.1577), but calibration improvement (3.4pp) falls below 5pp C1 threshold -> FAIL 4/5. C1 failure is partly attributable to nociceptive contamination in z_world: both split and unified z_world conditions still carry harm proximity signals that conflate calibration metrics. True calibration advantage of split should increase once SD-010 (separate z_harm) removes nociceptive contamination. Attribution advantage (82%) is real and unambiguous. SD-005 split is prerequisite for clean SD-010 implementation. EXQ-047d (2026-03-22, x2 runs): DESIGN FAILURE -- not a genuine negative. Probed frozen representations from training with no separation objective. Both z_self and z_world carried nearly equal action information (split_action_dissociation=0.008; need >0.10). Root cause: without adversarial pressure the encoder has no incentive to concentrate action info in z_self and expel it from z_world. The two runs are byte-identical (same seed, deterministic frozen probe) -- 1 effective data point, not 2. Evidence direction "weakens" on both runs is a tagging artefact of the design failure; these results should not be counted as genuine disconfirmation of SD-005. EXQ-047e FAIL (2026-03-23, 1/5): adversarial GRL on z_world. action_dissociation=0.055 (threshold 0.10 not met). Post-hoc GRL (80 episodes) too brittle; probe/encoder alternating update race is unstable. DESIGN FAILURE -- not a genuine disconfirmation. EXQ-047f FAIL x2 (2026-03-23, 2/5 each): orthogonal split with cosine penalty. action_dissociation=-0.108/-0.065 (NEGATIVE), orth_cos_mean_eval=0.002-0.018 (near zero). DESIGN FAILURE: geometric orthogonality is the WRONG test. Functional specialisation from differential loss routing does NOT require geometric orthogonality -- there are legitimate shared dimensions (top-down z_beta, blurry body-world boundary near hazards). Cosine constraint penalises both channels symmetrically, producing two decorrelated noisy representations rather than two specialised ones. EXQ-078 PASS (2026-03-23, 3/3): behavioural test -- split z_world produces gap_approach 0.197 vs unified 0.147 (delta=0.050 > 0.02 threshold). Split IS behaviourally better. This confirms SD-005 benefit at the functional level even though geometric probes failed. EXQ-047g QUEUED (2026-03-23): correct redesign -- functional cross-decoding probes. z_self -> action prediction acc vs z_world -> action prediction acc. z_world -> contact prediction acc vs z_self -> contact prediction acc. 400 warmup episodes (differential loss routing, no orthogonality constraint). Long run (~3h Mac). This is the proper test: does differential loss routing produce functionally specialised latents? Architecture note: E2.world_forward(z_world, action) -> z_world_next already implements an efference copy signal. Future architecture direction: use E2 prediction as the primary reafference signal (exafference = z_world_actual - E2(z_world_prev, action)), which would provide a stronger and self-improving separation mechanism. See SD-007 notes. depends_on: [] location: docs/architecture/sd_004_sd_005_encoder_codesign.md - id: SD-006 title: "E1, E2, and E3 run at characteristic rates (async multi-rate) rather than lockstep." claim_type: design_decision subject: control_plane.async_multirate_execution polarity: asserts status: implemented status_note: > E1, E2, E3 run at characteristic rates rather than lockstep (phase 1: time-multiplexed). Asynchronous multi-rate execution reflects distinct biological loop timescales (theta/gamma/beta bands). See ARC-023. evidence_quality_note: | EXQ-052 FAIL (2026-03-20): C3 failure was a diagnostic bug — experiment tracked len(agent.theta_buffer._buffer) where correct attribute is _z_world_buffer; AttributeError silently caught, max_theta_buffer_size=0 always. All other criteria passed: E3 functional at 9× slower rate (cal_gap=0.846), e3_tick_ratio=0.109 ≈ 1/9. EXQ-052b fixes the counter. Functionally, multi-rate execution is working. EXQ-052b PASS (2026-03-21, 5/5): ThetaBuffer attribute fix confirmed. max_theta_buffer_size correctly tracked via _z_world_buffer. Multi-rate execution fully functional: E3 runs at ~1/9 the step rate of E1, theta buffer fills and batches E1 outputs before each E3 sample. All five criteria met. SD-006 implementation validated end-to-end. depends_on: [ARC-023] location: docs/architecture/control_plane_heartbeat.md - id: SD-007 title: "ReafferencePredictor provides perspective-corrected world latent by subtracting self-caused sensory change." claim_type: design_decision subject: encoder.perspective_corrected_world_latent polarity: asserts status: implemented status_note: > ReafferencePredictor in ree_core/latent/stack.py: z_world_corrected = z_world_raw - ReafferencePredictor(z_world_raw_prev, a_prev). Input must be z_world_raw_prev (not z_self_prev): cell content entering view dominates Δz_world_raw and is inaccessible from body state alone (EXQ-027 run 1 R²=0.027 with z_self inputs). Biological basis: MSTd receives optic flow plus efference copy. Known limitation (Brooks & Cullen 2019 review): the brain continuously recalibrates the motor-sensory relationship; the current static predictor does not model this -- potential failure mode for non-stationary environments or long-running agents. depends_on: [SD-005] location: docs/architecture/sensory_stream_tags.md - id: SD-008 title: "LatentStack EMA alpha for z_world must be >= 0.9 to preserve event responsiveness." claim_type: design_decision subject: encoder.z_world_alpha_correction polarity: asserts status: stable status_note: > Promoted to stable 2026-03-30 (overall_conf=0.874, conflict_ratio=0, 5 supporting experiments, 2 lit entries, 0 weakening experiments in isolation). Promoted to provisional 2026-03-22. Two uncontested supporting experiments: EXQ-023 (alpha=1.0, event_selectivity_margin=0.084 passed SD-008 criterion; overall experiment failed on unrelated claims -- SD-007, ARC-016, SD-003 -- which have since been independently validated) and EXQ-040 (PASS 4/4, z-separation alpha=0.9 directly confirmed correct event responsiveness). Zero genuine weakening experiments: EXQ-023 direction corrected from weakens to supports; EXQ-024 and EXQ-025 had SD-008 removed per claim_ids accuracy rule (those experiments ran with alpha=0.9 but tested SD-003 pipeline, not whether alpha value matters). Evidence_direction systematic issue documented as known indexer limitation. Motivation: at alpha=0.3, z_world is a ~3-step weighted average suppressing event responses to 30%, trivialising E2_world prediction (MSE invariant to env perturbation) and preventing ARC-016 precision from firing. alpha_self may remain low (body state genuinely autocorrelated). Config: LatentStackConfig.alpha_world (default 0.3 for compat; set 0.9 or 1.0 for correct operation). EXQ-177 FAIL/mixed (integration test): event_selectivity_margin~0 in full-stack composition (SD-007 + SD-008 + SD-003) despite alpha_world=0.9. Attributed to destructive interference in combined stack, not SD-008 in isolation. Isolated SD-008 tests remain definitive; integration gap is an SD-003/SD-007 diagnostic target. depends_on: [SD-005] location: docs/architecture/sd_004_sd_005_encoder_codesign.md evidence_quality_note: | Promoted to stable 2026-03-30. overall_conf=0.874, conflict_ratio=0, 5 exp entries, 2 lit entries, 0 weakening experiments in isolation. Isolated SD-008 tests (EXQ-023, EXQ-040) are definitive and uncontested. EXQ-145 INCONCLUSIVE (2026-03-30): full-stack integration test; Phase 1 reafference gate failed (r2=0.076 < 0.08) before SD-008 selectivity was tested -- engineering bottleneck, does not challenge SD-008. EXQ-177 FAIL/mixed (2026-03-30): event_selectivity_margin~0 in full-stack (SD-007+SD-008+SD-003) despite alpha_world=0.9 and world_forward_r2=0.941 -- attributed to destructive interference in combined stack composition, not an isolated SD-008 failure. Integration stack diagnostic is a separate workstream (SD-003/SD-007 target). - id: SD-009 title: "z_world encoder requires event-type cross-entropy auxiliary loss during training." claim_type: design_decision subject: encoder.event_contrastive_supervision polarity: asserts status: provisional status_note: > Promoted candidate -> provisional 2026-04-03. Evidence: conf=0.768, conflict_ratio=0, 2 exp supports, 2 lit entries. EXQ-020 PASS (2026-03-20): event_classification_acc=0.692, selectivity_margin=0.882. MECH-100 (same claim, mechanism register) promoted to stable this session. SD-009 provisional: design decision validated, implementation confirmed in ree-v3 substrate. depends_on: [SD-005] location: docs/architecture/sd_004_sd_005_encoder_codesign.md - id: SD-010 title: "HARM stream is a dedicated sensory pathway separate from z_world processing." claim_type: design_decision subject: harm_stream.nociceptive_separation polarity: asserts status: implemented status_note: > The HARM stream (ARC-027) must be a separate sensory pathway independent of z_world: dedicated HarmEncoder(harm_obs → z_harm), not subject to reafference correction. E3.harm_eval takes z_harm as primary input. SD-003 attribution operates on harm stream output not full world latent. Resolves EXQ-027b over-correction, EXQ-043/044 calibration collapse, and EXQ-047 SD-005 calibration shortfall — all converge on fused z_world as root cause. Depends on ARC-027 implementation. evidence_quality_note: | EXQ-058b ARC-027 PASS (2026-03-20, 5/5 criteria): Foundational validation of the separate harm pathway. HarmEncoder trained with direct MSE supervision on hazard/ resource proximity labels achieves Pearson_r=0.973, calibration_gap=0.712, harm_pred_std=0.417. mean_harm_none=0.014 vs mean_harm_contact=0.987 — clean rank ordering across all event types. First V3 experiment showing z_harm can be learned without fusing into z_world. Confirms ARC-027 architectural separation is implementable; prerequisite for SD-003 + SD-010 joint pipeline tests. EXQ-056 FAIL (2026-03-20): Baseline HarmEncoder collapsed (z_harm=1.0 for all types, variance=0 to 14 d.p.). Root cause: HarmEncoder received saturated or constant harm_obs input; architecture was correct but input was not. EXQ-058b fixed by adding direct MSE supervision on proximity labels as the training signal. EXQ-058c post-fix FAIL (2026-03-20, 3/5 criteria): SD-003 + SD-010 joint test. calibration_gap=0.036 (threshold 0.05 not met); mean_harm_none=0.629 elevated because dense hazard environment ensures agent is nearly always near a hazard, inflating the "none" label harm estimate. C1/C2 fail; C3 (world_forward_r2=0.944) PASS; causal_sig_approach PASS but collapse guard fails. Dense environment calibration remains an open design issue. EXQ-056c PASS (2026-03-21, 4/5): SD-010 reafference isolation retest. Pearson r_z_harm=0.914. HarmEncoder correctly trained on harm_obs alone, not subject to reafference correction applied to z_world. Clean separation confirmed under varied experimental conditions. EXQ-059c PASS (2026-03-21, 4/5): MECH-102 via SD-010 pipeline. contact_rate_reduction of ~10× confirmed — agent proximity-avoiding behaviour measurably different with functional harm stream. Harm signal accessible during trajectory planning (not only at contact events), consistent with approach-gradient leverage predicted by MECH-102. depends_on: [SD-005, SD-007, ARC-027] location: docs/architecture/sensory_stream_tags.md - id: SD-011 title: "Harm stream separates into sensory-discriminative (z_harm_s: proximity/intensity, Adelta-pathway analog, forward-predictable) and affective-motivational (z_harm_a: accumulated homeostatic deviation, C-fiber/paleospinothalamic analog)." claim_type: design_decision subject: harm_stream.dual_nociceptive_streams polarity: asserts status: candidate implementation_phase: v3 claim_level: architectural depends_on: [SD-010, ARC-027] functional_restatement: > The single z_harm vector conflates two functionally distinct nociceptive signals: (1) z_harm_s: immediate proximity/intensity, analogous to the lateral spinothalamic tract (Adelta fibers, VPL -> S1/S2). Tracks stimulus with minimal lag; predictable from action (moving away from hazard reduces proximity predictably). This stream supports a learnable forward model E2_harm_s(z_harm_s, action) -> z_harm_s_next, enabling SD-003 counterfactual attribution within the harm stream. (2) z_harm_a: accumulated homeostatic deviation, analogous to the medial paleospinothalamic tract (C fibers, intralaminar thalamus -> ACC/insula/amygdala). Slower to change; integrative; represents sustained threat urgency and body state. NOT counterfactually modeled -- feeds E3 directly as motivational/urgency signal. ARC-016 harm variance gating operates on z_harm_a (motivational urgency scales with accumulated threat state), not z_harm_s. SD-003 counterfactual attribution operates on z_harm_s (via E2_harm_s forward model), not z_harm_a. evidence_quality_note: | EXQ-093 FAIL + EXQ-094 FAIL (2026-03-24): Both confirmed bridge_r2=0 -- SD-010 makes z_world perp z_harm by architectural design, so HarmBridge(z_world -> z_harm) is infeasible (nothing to learn). EXQ-094 confirmed training E3 on bridge noise produced 100x regression in harm_var vs EXQ-088. These experiments collectively demonstrate that the current single-stream z_harm cannot simultaneously serve SD-003 (needs forward model) and ARC-016 (needs sustained motivational signal). Stream separation is the solution confirmed by dual-pathway human nociceptive anatomy. Biological grounding: Melzack & Casey (1968) three-component pain model; Craig (2002, 2003, 2009) interoception and insular cortex homeostatic map; Rainville et al. (1997, Science) gold-standard dissociation: hypnotic modulation of unpleasantness modulates ACC but not S1 -- affective and sensory-discriminative components are independent. EXQ-198 PASS (2026-04-01): dual stream stability. dissociation_advantage=1.243, decision=retain_ree. harm_fwd_r2=0.677 (dual) vs 0.555 (fused). All seeds pass C1+C4. Manifest had status=UNKNOWN (old template) but metrics.json shows clear positive result. evidence_direction corrected to "supports". Confirms dual-stream separation is functionally advantageous over fused z_harm. Keltner et al. (2006) predictability suppresses sensory (S1) activity, not affective (ACC) -- forward model cancellation applies to z_harm_s but not z_harm_a. EXQ-100 FAIL 2/4 (2026-03-27): First integration of AffectiveHarmEncoder (z_harm_a). C3 PASS: step-std z_harm_a=0.026 vs z_harm_s=0.369 -- EMA smoothing structure confirmed at the encoder output level. C4 PASS (no errors). C1 FAIL: z_harm_a LOWER in HIGH_HAZARD (0.392) than LOW_HAZARD (0.425) -- inverted response. C2 FAIL: autocorr z_harm_a=-0.017 vs z_harm_s=0.061 -- negative autocorrelation from an EMA output is anomalous. Diagnosis: harm_obs_a (EMA accumulator in environment) may not vary across conditions as expected, OR AffectiveHarmEncoder is outputting near-random values despite structural smoothness. EXQ-100b FAIL 1/5 RAW_EMA_FAIL (2026-03-27): Two-phase diagnostic confirmed root cause is in CausalGridWorldV2, not the encoder. Raw harm_obs_a mean: HIGH_HAZARD=0.131 vs LOW_HAZARD=0.197 (inverted -- more hazards = lower accumulation). Raw autocorr_lag10=0.069 (expected ~0.37 for tau=20 EMA). Root cause identified: prox_now is normalized by hazard_max = field.max(). In high-density environments, overlapping hazard fields push hazard_max higher, shrinking each proximity contribution proportionally -- EMA tracks density-relative rather than absolute exposure. Fix: remove hazard_max normalization in prox_now computation (use raw field values or fixed normalization). EXQ-101 queued after fix. EXQ-106a PASS (2026-03-28): harm_obs_a persistence fix. evidence_direction=supports. Confirms harm_obs_a signal quality and accumulator behaviour are correct after normalization fix. Core z_harm_s/z_harm_a stream separation claim still not implemented; this validates the accumulator component only. EXQ-166a FAIL 4/6 (2026-03-30, shared result with SD-003/ARC-033): obs-space forward model approach cannot test z_harm_s stream directly. C2 causal gap near zero -- forward model learns general obs dynamics, not harm-stream-specific causal structure. SD-011 dual-stream architecture (z_harm_s forward-predictable) requires dedicated harm latent, not obs-space proxy. Evidence direction for this claim: mixed. EXQ-178b PASS 4/4 (2026-03-30): First validated PASS for SD-011 dual-stream dissociation. harm_fwd_r2=0.742 (C1: HarmForwardModel learns z_harm_s transitions). stream_corr=0.078 (C2: streams decorrelated, well below 0.85 saturation threshold). autocorr_gap=0.446 (C3: temporal integration confirmed -- z_harm_a accumulates vs z_harm_s). z_harm_s_hazard_corr=0.601 (C4: sensory stream encodes hazard proximity). EXQ-178/178a INCONCLUSIVE (C1 FAIL, identity collapse in earlier iterations, superseded). SD-011 stream separation validated; SD-003 redesign unblocked. EXQ-241 FAIL 1/3 criteria (2026-04-04): Diagnostic POC run. r2_s_forward=-0.32 (training/ wiring issue), harm_rate=1.0 in dual-stream conditions (POC wiring regression). Classified non-contributory (2026-04-06 governance): SD-011 stream separation already confirmed in EXQ-178b (r2=0.742) and EXQ-198 (dissociation_advantage=1.243). Diagnostic only. GOVERNANCE META (2026-04-06): All experimental evidence is diagnostic (EXQ-241 x2, EXQ-178 series). D3 reversal (R^2_affective > R^2_sensory in all seeds) is an active signal: AffectiveHarmEncoder is currently a monotone transform of the sensory stream, not a genuinely distinct pathway. Wiring needs a second source (context, history, residue field) inaccessible to the sensory stream. Status: pending_retest_after_substrate (gate: EXQ-247 wires z_harm_a into E3 commit gating, may create functional distinction). IMPLEMENTATION (2026-04-08): Second source implemented. AffectiveHarmEncoder now receives harm_history (FIFO rolling window of past harm_exposure scalars, length=harm_history_len) concatenated with harm_obs_a. Auxiliary harm_accum_head predicts accumulated harm scalar (MSE loss, weight=z_harm_a_aux_loss_weight). This gives z_harm_a genuinely distinct temporal information (harm trajectory over time) that z_harm_s does not receive, resolving the D3 reversal. Config: LatentStackConfig.harm_history_len (default 0, set 10 to enable). EXQ-241a validation queued. Status: ready for retest. notes: > Implementation in CausalGridWorldV2: emit harm_obs_a (EMA of harm_obs_s, tau=10-30 steps) alongside harm_obs_s (renamed from harm_obs). HarmEncoderA(harm_obs_a -> z_harm_a) added; HarmEncoder renamed HarmEncoderS. LatentState gains z_harm_a field. E3Selector takes z_harm_s (attribution) and z_harm_a (commit gating) as separate inputs. E2_harm_s forward model is a prerequisite (ARC-033). See thought intake evidence/planning/thought_intake_2026-03-24_dual_nociceptive_streams.md # Registered 2026-03-21 — Respiratory oscillator cluster # Source: cross-species behavioural observation (dog + human exhalation on plan abandonment) # Proposed by: Daniel Golden - id: MECH-107 title: "Exhalation is the physiological instantiation of E3 trajectory-abandonment — each exhale sweeps pre-commit rollouts for viability and clears the hypothesis tag." claim_type: mechanism_hypothesis subject: respiratory.exhalation_trajectory_abandonment polarity: asserts status: candidate implementation_phase: v3 claim_level: functional functional_restatement: > Each exhalation event triggers a sweep of currently active pre-commit hippocampal rollouts and DMN simulations for viability. Trajectories failing the sweep are abandoned and the hypothesis tag (MECH-094) cleared, opening the next deliberation cycle. In an ANN substrate, implemented as an explicit plan-sweep trigger fired at respiratory cadence, not as respiratory muscle activity. The biological mechanism (vagal afference from lung stretch receptors modulating brainstem–thalamic–prefrontal circuits) is the proposed substrate; the functional requirement is a periodic viability sweep co-occurring with the exhalation phase. depends_on: - MECH-094 - MECH-091 - ARC-023 location: docs/architecture/control_plane_heartbeat.md#mech-107 source: - docs/architecture/control_plane_heartbeat.md evidence_quality_note: | LIT-0094 (2026-03-23): Phase framing partially revised. Zelano 2016 (PMID 27927961): oscillatory power peaks during INHALATION, not exhalation; fear/memory performance 20-30% higher on inhalation. Boyadzhieva 2021 (PMID 34267621): inhalation = bottom-up sensory (feedforward); exhalation = top-down prediction propagation. Revised two-phase framing: inhalation = scan / viability evaluation (peak excitability, lower decision boundary); exhalation = execute abandonment / clear hypothesis tag. The claim captures the exhalation execution side; the scan phase on inhalation is implicit but now documented. Braendholt 2025 (PMID 40424351): inspiration lowers decision boundary (threshold channel) independently of valence bias channel. Vlemincx 2010 (PMID 20536901): sigh rate increases on load-recovery; sighs restore respiratory variability after sustained commitment -- direct behavioral evidence. notes: > Cross-species convergence: dogs exhibit a swift, audible exhalation precisely when abandoning a behavioural plan (e.g., approaching then breaking off); humans sigh in functionally identical contexts. Cross-species presence removes "cultural artifact" as an objection -- this is an evolutionarily conserved mechanism for plan-state management, not learned behaviour. Each exhalation is the physiological instantiation of the hypothesis-tag clear described in MECH-094: the body enacts what the planning system has decided (trajectory abandoned -> write gate cleared -> new deliberation cycle opens). Two-phase cycle (LIT-0094): inhalation = viability scan (peak neural excitability, reduced decision boundary per Braendholt 2025); exhalation = execute abandonment decision (top-down prediction propagation per Boyadzhieva 2021). Predicts: exhalation rate during deliberation tasks should correlate positively with plan-abandonment rate and negatively with committed-trajectory persistence. - id: MECH-108 title: "Respiratory rhythm (~0.2 Hz) is E3's plan-sweep oscillator — a third control-plane clock governing periodic sweeps of pre-commit hippocampal rollouts and DMN simulations." claim_type: mechanism_hypothesis subject: respiratory.breath_rate_plan_sweep_clock polarity: asserts status: candidate implementation_phase: v3 claim_level: functional functional_restatement: > The ~0.2 Hz respiratory cycle gates periodic sweeps of pre-commit hippocampal rollouts and DMN simulations for viability and abandonment. This is a third oscillator in the control-plane cluster, distinct from (a) heartbeat (~1 Hz, involuntary, arousal-level gating of E3 refresh rate — MECH-093) and (b) neural theta/beta (~4–30 Hz, sensory batching and commitment propagation — MECH-089, MECH-090). Semi-voluntary: top-down modulation is possible; cannot be fully stopped. That semi-voluntary property is exactly right for the planning layer — it is where deliberate top-down intervention in the planning cycle is possible. In an ANN substrate, implemented as a configurable sweep-clock parameter operating at ~0.2 Hz relative simulation time, orthogonal to the E3 heartbeat parameter. depends_on: - ARC-023 - MECH-107 - MECH-092 - MECH-093 - MECH-104 location: docs/architecture/control_plane_heartbeat.md#mech-108 source: - docs/architecture/control_plane_heartbeat.md notes: > LIT-0094 (2026-03-23): Empirical grounding confirmed. Zelano 2016 (PMID 27927961): human respiratory 0.16-0.33 Hz; system-wide entrainment of piriform cortex, amygdala, hippocampus. Tort 2018 (PMID 29691421): RCOs in all 15 brain regions simultaneously including nucleus accumbens and mPFC -- global clock. Yackle 2017 (PMID 28360327): preBotC projects monosynaptically to LC -- respiratory pacemaker is hardwired to NE precision modulation (MECH-104 direct dependency added). Braendholt 2025 (PMID 40424351): inspiration lowers decision boundary AND shifts valence bias -- two independent computational channels, not one scalar. Neuromodulator duality (Lottem 2018 PMID 29520000, Dayan & Yu 2006 PMID 17162459): inter-sweep phase = 5-HT-dominant (persist/commit); sweep phase = NE-ready (interrupt/ release). BreathOscillator approximates oscillation between these neuromodulator states. Three-oscillator control-plane hierarchy (slow to fast): (1) Respiratory plan-sweep (~0.2 Hz, semi-voluntary) -- sweeps pre-commit rollouts for viability/abandonment; inhalation = scan phase, exhalation = execute phase. (2) E3 heartbeat (involuntary thalamic pacemaking, ~0.5-2 Hz depending on z_beta arousal, MECH-093) -- governs E3 refresh rate and harm attribution resolution. (3) Neural theta/beta (~4-30 Hz, MECH-089, MECH-090) -- sensory batching and commitment propagation gating. Infraslow oscillations (Monto et al. 2008, PMID 18701689, 0.01-0.1 Hz) provide a fourth slower layer -- 55% detection probability modulation by ISF phase. Relevant to long-timescale MECH-047 hysteresis. Not modelled in BreathOscillator but noted. Quiescent replay (MECH-092) operates below all four, during genuine idle periods. - id: MECH-109 title: "Voluntary respiratory modulation provides the one top-down deliberate handle on the E3 plan-sweep clock; breath-hold is a predicted failure mode suppressing plan abandonment under anxious concentration." claim_type: mechanism_hypothesis subject: respiratory.voluntary_breath_control_planning_gate polarity: asserts status: candidate implementation_phase: v3 claim_level: functional functional_restatement: > Breathing is the only control-plane oscillator with a genuine top-down voluntary handle: the agent can deliberately sigh (force a plan-sweep cycle), slow breathing (reduce sweep rate), or hold breath (suppress plan sweeps entirely). This voluntary controllability — absent from heartbeat and from neural theta/beta — is structurally appropriate for the planning layer because deliberate plan abandonment requires an agent-initiated trigger. In an ANN substrate, the "voluntary respiratory modulation" is an explicit external trigger API that forces a plan-sweep cycle on demand, independently of the background respiratory cadence. depends_on: - MECH-108 - MECH-091 - MECH-107 evidence_quality_note: | LIT-0094 (2026-03-23): First direct behavioral evidence. Vlemincx et al. 2010 (PMID 20536901, Psychophysiology): sigh rate increased during and immediately after mental load / sustained attention tasks. Sustained attention compressed total respiratory variability (reduced flexibility). Sighs functioned as resets -- restoring respiratory variability after high-commitment compression. This is direct empirical support: high-commitment states suppress spontaneous plan-sweep cycling (compressed variability); sighs are the voluntary restoration event (reset). Timing: sigh rate elevated on task-offset / load-recovery, matching the prediction that sweep-clock suppression under load is followed by burst of sweeps at release. Consider upgrading to provisional after one targeted sigh-latency experiment. location: docs/architecture/control_plane_heartbeat.md#mech-109 source: - docs/architecture/control_plane_heartbeat.md notes: > Predicted failure mode: anxious breath-holding. Under high-stakes deliberation, agents (human and animal) frequently hold their breath. This suppresses the respiratory plan-sweep clock (MECH-108) at exactly the moment the planning system is under maximal load and most needs plan-abandonment sweep capacity. The failure is self-compounding: suppressed sweeps prevent abandonment of non-viable trajectories, increasing deliberative load, increasing anxiety, sustaining the breath-hold. This also coincides with the agent needing more oxygen (high metabolic demand from high arousal), creating a somatic conflict between gas-exchange need and plan-sweep suppression. Clinical implication: deliberate breathing (respiratory rate restoration) is not merely calming through vagal activation but specifically restores the plan-abandonment sweep cycle. "Taking a deep breath" is mechanistically plan-sweep restoration. Voluntary sigh = deliberately forced single-cycle abandonment sweep. Contrast: heartbeat cannot be voluntarily overridden (modulated indirectly by arousal and vagal tone, not directly commanded); neural theta/beta are wholly involuntary. Respiratory rate is the unique locus of deliberate top-down planning-clock intervention. - id: MECH-110 title: "Laughter is rapid repeated hypothesis-tag cycling — threat activation followed immediately by safe-resolution — predicting incongruity+resolution as the necessary structure of humour." claim_type: mechanism_hypothesis subject: respiratory.laughter_rapid_tag_cycling polarity: asserts status: candidate implementation_phase: v3 claim_level: functional functional_restatement: > Laughter consists of repeated forced exhalations, each constituting one plan-sweep cycle (MECH-107). A laughter episode is therefore rapid repeated oscillation through: threat-hypothesis activation (E3 writes a harm-evaluation trajectory tag) → sweep → safe-resolution (sweep finds no genuine harm → tag cleared via exhalation). The pleasure of laughter is the affective signature of rapid repeated threat-non-confirmation. This predicts that humour requires (1) an initial threat/incongruity activation (hypothesis tag written) and (2) a fast safe-resolution (sweep clears tag). Without the initial activation, no laughter; without fast resolution, distress rather than laughter. This matches incongruity-resolution theories of humour independently. Social extension (V5-scoped): synchronised group laughter is a shared respiratory phase-reset signal synchronising E3 plan-sweep clocks across agents, functioning as a joint "threat absent" signal. Contagious laughter is clock entrainment across agents. depends_on: - MECH-107 - MECH-094 - MECH-091 - MECH-108 location: docs/architecture/control_plane_heartbeat.md#mech-110 source: - docs/architecture/control_plane_heartbeat.md notes: > Each forced exhalation in laughter is a rapid hypothesis-tag clear (MECH-107). The density of exhalations in laughter (typically 3–7 per second) far exceeds the resting respiratory rate (~0.2 Hz), indicating that laughter is a deliberate high-frequency override of the background plan-sweep clock. This high-frequency override is tolerated because each tag-clear is immediately followed by re-activation and re-clearing: the threat level perceived to be genuinely zero. The repeated cycling is itself the mechanism of pleasure — not a single resolution but the kinetic experience of rapid iterated safe-confirmation. Distinguishes: (a) single exhalation/sigh = plan abandonment under genuine uncertainty; (b) laughter = rapid safe-confirmation cycling where genuine threat is absent. Predicts: humour that initially activates a stronger threat tag should produce more intense laughter when resolved safely (gallows humour, dark comedy). Predicts: incomplete resolution (ambiguous punchline) produces nervous laughter or discomfort rather than clean laughter — the tag-clear is partial and the residue persists. Social synchronisation (V5-scoped): group laughter as joint respiratory entrainment synchronises E3 plan-sweep clocks across agents, functioning as a distributed "nothing to fear" signal. Contagious laughter is clock entrainment. Social laughter evolution may be grounded in this synchronisation function rather than (or in addition to) pure affiliation signalling. # Registered 2026-03-22 -- approach/avoidance drive symmetry cluster - id: MECH-111 title: "E1 prediction error surprise at moderate magnitudes generates intrinsic positive valence (curiosity/novelty drive); information-seeking is architecturally grounded." claim_type: mechanism_hypothesis subject: drive.epistemic_value_novelty polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | EXQ-141 FAIL/weakens (2026-03-29): novelty drive discriminative pair FAIL. First experimental entry. depends_on: - MECH-069 - ARC-030 location: docs/architecture/approach_avoidance_symmetry.md#mech-111 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > Computational analogue of FEP epistemic value (expected information gain), distinct from pragmatic value (harm reduction). E1 surprise is not only alarming -- at moderate magnitudes it is intrinsically rewarding. Without this, the agent has no intrinsic motivation to encounter novel states; exploration decays to zero under harm-avoidance pressure alone. The architecture needs to distinguish alarm-surprise (high magnitude, harmful context) from curiosity-surprise (moderate magnitude, safe novel context). This distinction may route through z_beta. Registered 2026-03-22. Literature: Ogasawara et al. 2022 (Nat Neurosci, DOI 10.1038/s41593-021-00950-1) -- novelty seeking has its own dedicated neural circuit (temporal cortex-zona incerta pathway) DISTINCT from reward dopamine; ZI neurons activate on predictions of novelty before action. Wang et al. 2024 (PLoS Comput Biol, DOI 10.1371/journal.pcbi.1011516) -- dopaminergic novelty signals encode VARIANCE of reward distribution (uncertainty), not mean reward; distinct from TD prediction error. Kakade & Dayan 2002 (Neural Netw, DOI 10.1016/s0893-6080(02)00048-5) -- dopamine multiplexes exploration bonuses alongside reward prediction error; novelty bonuses lower commit threshold for novel states. Suri et al. 2001 (Neuroscience, DOI 10.1016/s0306-4522(00)00554-6) -- ablating novelty responses impairs planning, not just exploration. Architecture shape: novelty drive is NOT repurposed E1 prediction error; it encodes reward *variance* (uncertainty), needs own latent channel or modulator. Acts on commit threshold (lowering for novel states) rather than through goal attractor channel. Full synthesis: evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md - id: MECH-112 title: "E3 requires a structured latent goal representation (positive attractor in z_world or z_goal sub-space) distinct from harm avoidance." claim_type: mechanism_hypothesis subject: drive.goal_state_attractor polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-005 - ARC-007 - MECH-069 - ARC-030 location: docs/architecture/approach_avoidance_symmetry.md#mech-112 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > E3 error signal is specified as harm/goal error (MECH-069) but goal has no structural representation -- subsumed under harm with no latent encoding, no goal encoder, no approach gradient. The hippocampal planner must navigate toward attractors as well as away from repellers. Without goal attractors the residue field is a pure aversion landscape; gradient minimum under learning is quiescence. Goal representation needs its own latent (or sub-component of z_world), its own encoding path, and its own contribution to hippocampal terrain. INV-029 (love as long-horizon care-investment) and ARC-026 cash out architecturally here: coherence and flourishing are positive-valence attractors in z_world space. Registered 2026-03-22. Literature: Barch & Dowd 2010 (Schizophr Bull, DOI 10.1093/schbul/sbq068) -- critical distinction between LIKING (hedonic response at goal receipt) and WANTING (prospective motivational drive toward future states). Schizophrenia negative symptoms = impaired wanting with intact liking. REE may have liking (z_beta valence) but lacks wanting (goal representation). MECH-112 must be prospective and persistent, not a terminal reward signal. Culbreth et al. 2023 (Psychol Med, DOI 10.1017/S0033291722003993) -- reduced reward- seeking correlated with loss-avoidance bias in schizophrenia; rlPFC encodes uncertainty- driven exploration. Hollerman et al. 2000 (Prog Brain Res, DOI 10.1016/S0079-6123(00)26015-9) -- striatum provides neural representation of goal; subgoal representations at multiple trajectory points, not only terminal state. Rusu & Pennartz 2019 (Hippocampus, DOI 10.1002/hipo.23167) -- hierarchical BG loops; limbic-affective loop sets goal context for dorsomedial PFC-striatal loop; hippocampal SWR replay for goal consolidation. Architecture shape: MECH-112 = "wanting" system; prospective, persistent goal state in E3 domain; hippocampal terrain becomes attractive near z_goal, not just repulsive near z_harm; subgoal representations needed at multiple trajectory steps. Full synthesis: evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md evidence_quality_note: | EXQ-074c superseded (2026-03-27): resource_respawn bug -- zero resource visits in all conditions. Superseded by EXQ-074d. EXQ-074d FAIL 3/4, EXQ-074e FAIL 3/4 (2026-03-27): C1 FAIL in both (resource_rate_gap=0). C1 confound: greedy navigation applied uniformly to all conditions -- wanting cannot show a behavioural lift above nogo even when z_goal is active (goal_active=True, goal_norm=0.28 in 074d). C2/C3/C4 PASS in both: wanting/liking dissociation confirmed in trajectory scoring metrics (l2_redirect, l1_fraction, goal_proximity). C1 requires redesign: use goal-following navigation in wanting condition vs random/nogo baseline, so resource visit rate CAN vary across conditions. EXQ-085g FAIL 3/4 (2026-03-29): same C1 failure pattern. goal_resource_r=0.066 < 0.2 threshold: z_goal not pointing toward resources despite contact-gated seeding. z_world at contact encodes full scene; resource features not isolated. SD-015 z_resource separation required before MECH-112 wanting system can navigate to goals. EXQ-085h through 085l FAIL C2 (2026-03-30): SD-015 encoder series (5 iterations, 085l final, supersedes prior). 085l: prox_r2=0.908 (z_world encodes proximity), goal_resource_r_enc=0.869 (ResourceEncoder learns position-invariant features), but benefit_ratio=0.420 (navigation not improved). The signal is learnable; the bottleneck is action-selection integration. Governance: hold -- promotion blocked by consistent C2 FAIL across all 5 iterations. MECH-112 wanting system requires working navigation (C2) not just representation (C1/C3) to validate the claim. Goal-lift battery (2026-04-01): EXQ-185 (direct prox-argmax, bypass z_goal entirely), EXQ-186 (hybrid benefit+harm selector), EXQ-183 (BCE shared selector) -- ALL FAIL. Systematic pattern: every learned 1-step greedy mechanism performs WORSE than random (0.21x, 0.25x, 0.73x respectively). EXQ-186 4-condition decomposition: random=0.862 benefit/ep, benefit_only=0.350, combined=0.212, harm_only=0.137. Directed mechanisms increase harm 2-2.3x relative to random while collecting fewer resources. Diagnosis: 1-step greedy action selection creates local traps near hazards; random walk's advantage is exploration breadth on 10x10 grid with 4 respawning resources. CRITICAL GATE: EXQ-182a (oracle ceiling) queued -- if perfect-information 1-step greedy also fails, the env config is near-optimal for random walk and no learned mechanism can beat it without multi-step planning (SD-004 hippocampal navigation). EXQ-074f PASS 4/4 (2026-04-04): FIRST full PASS for MECH-112 wanting system with correct navigation design. wanting l1_fraction=0.96 (agent persistently approaches L1 post-relocation); liking l1_fraction=0.43; resource_rate=0.148 (floor PASS >= 0.05). C4 dissociation confirmed (delta=0.53 > 0.10). Prior design confound resolved: wanting condition uses spatial greedy toward L1; nogo/liking use standard action selection. Single seed only (42); prior 085 series FAIL history means multi-seed replication still needed before promotion. Navigation integration works at small scale (6x6 grid); larger grid performance (10x10, SD-004 planning) remains to be tested. EXQ-074e superseded (2026-04-04): same design confound as 074d (uniform greedy). 074f FAILs 1-3 superseded: exploration iterations before final criteria/navigation design. EXQ-225 FAIL non_contributory (2026-04-04): goal-lift redesign (GOAL_PROXIMATE vs GOAL_AGNOSTIC CEM scoring). resource_lift=-0.013 (COMBINED slightly worse). z_goal seeding confirmed (avg_goal_norm=0.383). This FAIL is not evidence against MECH-112 wanting architecture. It is an action-selection bottleneck: on a 10x10 grid with 4 respawning resources, 1-step greedy CEM consistently underperforms random walk (same pattern as EXQ-183/185/186 battery). The wanting signal is present and active; the mechanism that fails is translation of wanting into resource lift without multi-step planning (SD-004). Wanting representation (EXQ-074f PASS) and wanting/liking dissociation are validated; wanting -> resource collection requires hippocampal navigation. evidence_direction corrected to non_contributory in manifest. EXQ-233 did not provide evidence for MECH-112: z_goal seeding never fired in either condition due to sub-threshold benefit_exposure (proximity_benefit_scale=0.03 default, benefit_exposure~0.025 < threshold 0.1). EXQ-238 with corrected env params will provide the clean causal ablation. EXQ-234 PASS (2026-04-05): independent replication of wanting/liking dissociation with seeds [1,2,3]. All 4 criteria met in all 3 seeds. Replicates EXQ-074f. GOVERNANCE META (2026-04-06): Illusory conflict resolution risk. Supports come exclusively from wanting/liking circuit (EXQ-074f, EXQ-234) which tests incentive salience -- a narrow z_goal pathway. The broader claim (structured goal latent for flexible goal-directed behavior) is untested. Non-contributory FAILs (EXQ-235) removed from scoring but indicate z_goal alignment failure persists outside the wanting pathway. Status: pending_retest_after_substrate (gate: EXQ-247 SD-011/SD-012 integration). - id: MECH-113 title: "The agent requires a self-maintenance error signal monitoring internal latent coherence, independent of external harm signals." claim_type: mechanism_hypothesis subject: drive.homeostatic_self_coherence polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | EXQ-142 FAIL/weakens (2026-03-28): self-maintenance pair FAIL (first run). EXQ-142 FAIL (2026-03-29): self-maintenance pair re-run, FAIL again. Two consistent FAIL results. Weakening evidence accumulating. depends_on: - SD-005 - SD-007 - INV-030 - ARC-030 location: docs/architecture/approach_avoidance_symmetry.md#mech-113 source: - evidence/planning/thought_intake_2026-03-22_approach_avoidance_drives.md notes: > Degradation of z_self stability, prediction fidelity, or reafference loop integrity is undetected by harm-avoidance machinery if no external harm is occurring -- agent can fail silently. A homeostatic coherence signal closes this gap and provides intrinsic motivation for continued self-regulation. Implements INV-030 (viability as binding constraint) at the architectural level. Two framings require separate experimental testing: (1) z_self coherence signal -- monitor z_self statistical properties (entropy, prediction accuracy, drift) and generate a self-maintenance error when coherence degrades below threshold; (2) FEP Markov blanket resistance -- self-maintenance emerges from minimising surprise about self-generated states, framing self-coherence as a prior over blanket states rather than an explicit signal. Both framings are architecturally open; both need testing. pending_design: framing (D_eff vs. Hopfield stability vs. Markov blanket resistance -- see Q-022 dissociation question). Registered 2026-03-22, updated 2026-03-23. Literature: Seth & Friston 2016 (Phil Trans R Soc B, DOI 10.1098/rstb.2016.0007) -- interoceptive inference as inversion of a generative model of internal milieu; sense of self emerges from maintaining the model. Markov blanket framing and z_self coherence signal framing converge: maintaining z_self fidelity IS maintaining the Markov blanket. Petzschner et al. 2021 (Trends Neurosci, DOI 10.1016/j.tins.2020.09.012) -- three levels: homeostatic (reactive), allostatic (predictive setpoint), goal-directed. All three levels correspond to distinct generative models in active inference; may need all three in REE. Stephan et al. 2016 (Front Hum Neurosci, DOI 10.3389/fnhum.2016.00550) -- CRITICAL: a third level beyond the two framings -- ALLOSTATIC SELF-EFFICACY, metacognitive monitoring of the agent's own regulatory capacity. Chronic dyshomeostasis + low self-efficacy = learned helplessness. Direct mechanistic link: MECH-113 failure -> low self-efficacy -> behavioral inhibition -> Q-021 behavioral flatness. Barrett et al. 2016 (Phil Trans R Soc B, DOI 10.1098/rstb.2016.0011) -- depression = locked-in brain insensitive to sensory context = clinical manifestation of MECH-113 failure. Architecture shape: three-level implementation (updated 2026-03-23): Level 1 (reactive): z_self D_eff monitoring + self-maintenance loss. IMPLEMENTED in agent.py (compute_z_self_d_eff, compute_self_maintenance_loss). TESTED by EXQ-075. Level 2 (allostatic anticipatory): E2 monitors predicted z_self D_eff over rollout horizon and generates anticipatory setpoint signal -- "I am heading toward self-model incoherence in N steps." NOT IMPLEMENTED. GATED BY ARC-031 (V4): Level 2 requires hippocampal z_self trajectory navigation to plan deliberation sequences that maintain low D_eff. Do not attempt Level 2 experiments until ARC-031 waking-architecture prerequisites (MECH-113/114 EXQ-075/076 PASS) are met and V4 substrate exists. Level 3 (metacognitive self-efficacy): E3 models its own regulatory capacity; chronic self-incoherence + low regulatory belief = learned helplessness -> Q-021 Pathway B. Level 3 is also V4 scope; partially approximated by MECH-114 D_eff commit gate, but the full metacognitive framing requires richer self-model machinery. D_eff framing (epistemic-mapping): D_eff = (sum|z_self|)^2 / sum(z_self^2) -- participation ratio measuring effective dimensionality. Implemented in EXQ-075. Hopfield stability framing: familiarity of z_self patterns relative to stored LRU memory. NOT tested. Creates dependent claims MECH-116 (familiarity signal) and MECH-117 (coherent-but-unfamiliar pathology = low D_eff + low stability, distinct from high D_eff dispersion). See Q-022 (dissociation question: can D_eff and Hopfield stability dissociate?). Markov blanket framing: self-maintenance as minimising surprise about self-generated states (FEP formulation). Unoperationalised as of 2026-03-23. Dissociation question: can an agent have low D_eff (coherent) but high FEP self-surprise (predictable self-model but surprising self-generated states)? If yes, Markov blanket and D_eff framings are testing different things and MECH-113 should eventually be split. Dissociation test registered as EVB-0069. EXQ-075 PASS is necessary but not sufficient for MECH-113: it confirms Level 1 D_eff reactive homeostasis only. Levels 2 and 3 and the Hopfield/Markov framings remain open. Full synthesis: evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md - id: MECH-114 title: "D_eff (z_self participation ratio) is a necessary co-condition for commitment, alongside world-side running_variance; both must be in range before irreversible action." claim_type: mechanism_hypothesis subject: commit_gate.z_self_d_eff_gate polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-016 - MECH-113 - INV-033 - SD-005 location: docs/architecture/approach_avoidance_symmetry.md#mech-114 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md notes: > Current ARC-016 commit condition: running_variance < commit_threshold (world-side confidence). MECH-114 extends to an AND condition: running_variance < commit_threshold AND d_eff < d_eff_threshold (self-side coherence). These are independent signals -- an agent can have accurate world predictions (low running_variance) while having an incoherent self-model (high D_eff). MECH-114 prevents commitment in this dissociated case. Biological motivation: decision to commit to irreversible action requires both knowledge of what will happen (world confidence) and knowledge of what one is doing (self-coherence). Acting without self-coherence is analogous to acting under dissociation -- the agent cannot be held responsible for an action it did not, in some sense, knowingly perform. Threshold calibration: EXQ-075 provides baseline D_eff values (pre-perturbation ~1.5x target) that can be used to set d_eff_threshold. A reasonable initial value is d_eff_threshold = d_eff_baseline * 1.5 (same as MECH-113 homeostatic target). Implementation: E3TrajectorySelector.select() checks both running_variance and current d_eff before setting committed=True. REEAgent.compute_z_self_d_eff() provides the D_eff value at each step. EXQ-075 is the prerequisite: must confirm D_eff monitoring works before wiring into commit gate. Follow-on experiment: EXQ-076 (MECH-114 D_eff commit gating -- does D_eff gate reduce false commitment under self-model perturbation?). pending_design: threshold value and how to expose d_eff to E3 commit logic (currently computed in REEAgent but not passed to E3TrajectorySelector.select()). Registered 2026-03-23. - id: MECH-115 title: "Hypothesis tag (MECH-094) reliability degrades with z_self dispersion; high D_eff blurs the simulation/real-action boundary at the self-model level." claim_type: mechanism_hypothesis subject: hypothesis_tag.z_self_coherence_dependency polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-094 - MECH-113 - MECH-114 - INV-033 location: docs/architecture/approach_avoidance_symmetry.md#mech-115 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md notes: > MECH-094 (hypothesis tag = categorical phi(z) write gate) distinguishes simulated from real-action content. The tag is issued by the agent to mark replay/simulation as hypothesis-tagged, preventing residue accumulation from simulated harm. MECH-115 identifies a failure mode distinct from MECH-094 tag LOSS (the severe failure modeled as PTSD/psychosis mechanism): tag DEGRADATION due to self-model incoherence. When z_self is dispersed (high D_eff), the agent cannot reliably distinguish "I am simulating X" from "I am doing X" at the self-model level. The tag is still issued, but its reliability as a separating signal degrades. Mechanistic picture: the hypothesis tag is written by a gate conditioned on z_self state. If z_self is high-D_eff (diffuse across many dimensions), the gate's input representation does not clearly encode the current mode of operation. The tag becomes a noisy signal -- not absent, but unreliable. This is graded, not binary. Clinical analog: dissociative states where the patient performs actions and then reports uncertainty about whether they "really did" those things. Not full psychosis (MECH-094 tag loss), but a softening of the simulation/real boundary. Connection to INV-033: MECH-115 shows WHY second-order self-monitoring (INV-033) is necessary -- without it, MECH-094 cannot reliably perform its core function. The hypothesis tag is not just a write gate; it is a gate that depends on self-coherence. pending_design: how to implement graded tag reliability (precision weighting of tag signal? confidence score appended to hypothesis_tag field? downstream discounting of residue accumulation from uncertain-tag events?). Distinct from current hypothesis_tag: bool -- may require hypothesis_tag: float [0,1] in a future version. Registered 2026-03-23. - id: MECH-118 title: "Hopfield-style pattern familiarity of z_self states (stability against stored self-state memories) is a distinct self-maintenance signal from D_eff coherence." claim_type: mechanism_hypothesis subject: self_maintenance.hopfield_familiarity_signal polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | EXQ-143 FAIL/mixed (2026-03-29): Hopfield familiarity discriminative pair FAIL. First experimental entry. EXQ-084d SUPPORTS (per-claim override, 2026-03-30): stability dissociation confirmed -- stab collapses with both noise (0.063 vs 0.950) and novelty (0.026 vs 0.950) perturbation. Stability signal responds to perturbation as distinct signal from D_eff (D_eff remains flat ~20.6-20.8 across R1/R2). Supports that Hopfield-style familiarity (stability) is a discriminable self-maintenance signal. D_eff direction between regimes remains unresolved. depends_on: - MECH-113 - INV-033 - SD-005 location: docs/architecture/approach_avoidance_symmetry.md#mech-118 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md - evidence/planning/literature_synthesis_2026-03-23_self_monitoring_dissociation.md notes: > D_eff (MECH-113) measures coherence -- how focused z_self is across its dimensions. MECH-116 adds a complementary signal: familiarity -- whether the current z_self pattern matches any previously stored self-state. Both signals are needed for full self-monitoring because they can dissociate (see Q-022): a focused (low D_eff) self-model can be in an unfamiliar configuration, and a dispersed (high D_eff) self-model can resemble past dispersed states. Biological analog: hippocampal pattern completion (familiarity/recognition) operates independently of representational coherence (binding quality). Literature support: Yonelinas 2001 (PMID 11571028, DOI 10.1098/rstb.2001.0939) canonical dual-process review -- familiarity (signal-detection, perirhinal cortex) and recollection (threshold, hippocampus-dependent) are doubly dissociable. Maps to MECH-116 Hopfield familiarity vs MECH-113 D_eff. Aggleton et al. 2005 (PMID 16154457, DOI 10.1016/j.neuropsychologia.2005.01.019): patient KN -- 45% bilateral hippocampal volume loss, near-normal perirhinal cortex; near-normal familiarity, severely impaired recollection. Clearest existence proof that familiarity survives without the hippocampal substrate. Staresina et al. 2012 (PMID 22751037, DOI 10.1038/nn.3154): simultaneous fMRI + intracranial EEG -- familiarity (perirhinal) and recollection (hippocampal) process on distinct time bases, providing a temporal dissociation testable in z_self monitoring timecourses. D_eff and familiarity are orthogonal by construction: Rigotti et al. 2013 (PMID 23685452, DOI 10.1038/nature12160) -- participation ratio measures capacity/ambient dimensionality, not proximity to a stored attractor. High D_eff + no familiar states (novel high-D space) and low D_eff + very familiar states (settled attractor) are both possible. Implementation source: epistemic-mapping repo (dgolden), Epistemic_monitor.py. HopfieldMemory class: 64-slot LRU memory, modern Hopfield network (Ramsauer et al. 2021), attention-based retrieval, stability = max(softmax(beta * similarities)). At each step: store z_self; retrieve similarity to stored set; high stability = familiar self-state. Combined certainty (from epistemic-mapping): 0.4*(1-entropy/10) + 0.3*(1-D_eff/n) + 0.3*stability. This multi-framing combined score is a candidate INV-033 implementation that is sensitive to both coherence (D_eff) and familiarity (stability) failure modes. pending_design: integration with REEAgent -- HopfieldMemory would need to be instantiated in agent.py and updated at each z_self step. Memory slots are LRU-managed; window size is a hyperparameter. Registered 2026-03-23. - id: MECH-119 title: "Low D_eff combined with low Hopfield stability (coherent but unfamiliar self-state) is a distinct pathological regime from high-D_eff dispersion." claim_type: mechanism_hypothesis subject: self_maintenance.coherent_unfamiliar_pathology polarity: asserts status: stable status_note: | Promoted provisional -> stable 2026-04-03. Evidence: conf=0.857, conflict_ratio=0, 6 supports + 1 mixed, 3 lit entries. EXQ-144 PASS (2026-03-29, 5/5) confirmed core dissociation. EXQ-084d per-claim override (2026-03-30 + 2026-04-02 runs): MECH-119=mixed. Stability dissociation confirmed (C2/C4 PASS) but d_eff prediction partially fails. The stability prediction (coherent-unfamiliar -> low stability) is robustly validated. The d_eff prediction requires further disambiguation. Stable on stability arm. implementation_phase: v3 evidence_quality_note: | EXQ-144 PASS (2026-03-29, 5/5): coherent-unfamiliar discriminative pair. All 5 criteria met. evidence_direction=supports. 0 conflict ratio; conf=0.81 with 2 experiment entries (both supports) and 3 literature entries. Clean promotion to provisional per governance decision 2026-03-29. EXQ-084d MIXED (per-claim override, 2026-03-30 + 2026-04-02 replication): stability dissociation confirmed (C2/C4 PASS -- both noise and novelty collapse stability from 0.950 to 0.063/0.026). However R3 (novelty/structured perturbation) produces d_eff=22.6 -- HIGHER than R2 noise (20.8) and R1 normal (20.6). MECH-119 predicts coherent-unfamiliar regime has LOW d_eff + low stability. The low-stability prediction is confirmed; the low-d_eff prediction is not. R3 novelty increases dimensionality rather than compressing to a coherent subspace. Promoted to stable on stability arm; d_eff arm requires further disambiguation (see Q-022). depends_on: - MECH-113 - MECH-118 - Q-022 location: docs/architecture/approach_avoidance_symmetry.md#mech-119 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md - evidence/planning/literature_synthesis_2026-03-23_self_monitoring_dissociation.md notes: > D_eff monitoring alone is insufficient because it cannot detect the coherent-but-unfamiliar regime: an agent with low D_eff (focused self-representation) and low Hopfield stability (no stored self-state resembling current z_self) is in a state that appears confident by D_eff metrics but is epistemically dangerous because the agent has no prior experience to calibrate against. This is the hyperarousal analog: high apparent self-confidence but in uncharted internal territory. Clinical signature: depersonalisation/derealisation with preserved cognitive clarity. The patient can think clearly (coherent self-model, low entropy) but does not recognise their own mental states as familiar (low pattern familiarity). Distinct from confusion (high D_eff, dispersed self-model) and distinct from normal operation (low D_eff, high familiarity). Three regimes form a diagnostic space: (1) Normal: low D_eff, high stability (coherent + familiar) (2) Dispersed: high D_eff, low stability (confused -- classic MECH-113 failure) (3) Hyperarousal: low D_eff, low stability (coherent but in novel territory) Regime 3 is undetectable by D_eff-only monitoring. A self-maintenance system that only penalises high D_eff would miss Regime 3 entirely and might even reward the agent for being "coherent" while it is in a pathological state. Mechanistic consequence: the self-maintenance loss (MECH-113) and the commit gate (MECH-114) should incorporate BOTH D_eff and Hopfield stability to span all three regimes. Clinical literature support: Sierra & David 2011 (PMID 21087873, DOI 10.1016/j.concog.2010.10.018) -- depersonalization = increased prefrontal inhibition of insula/limbic. Self-representation structurally intact (low D_eff, coherent), emotional familiarity signal absent (low Hopfield stability), Markov blanket intact. This is the existence proof that Regime 3 (coherent + unfamiliar) occurs clinically. Paul et al. 2019 (PMID 31103548, DOI 10.1016/j.bpsc.2019.03.007): reduced EBA-to-DMN connectivity predicts depersonalization in depression -- body familiarity signal (EBA) severed from self-model integrity (DMN). Harricharan et al. 2017 (PMID 28911803, DOI 10.1016/j.neuropsychologia.2017.09.010): PTSD dissociative subtype -- self-model tightly organised around threat (low D_eff on danger dimensions, intact blanket) but body familiarity suppressed (perirhinal-interoceptive decoupling). All three framings dissociate simultaneously. Milliere 2017 (PMID 28588463, DOI 10.3389/fnhum.2017.00245): drug-induced ego dissolution -- dissociatives produce "coherent but detached" (low D_eff, low familiarity, disrupted blanket); psychedelics produce high D_eff with novel unfamiliar states; kappa-opioid agonists produce dysphoric unfamiliarity with intact self-location. Three drug classes simultaneously dissociate all three REE self-monitoring framings. Sierra et al. 2002 (PMID 11909918, DOI 10.1136/jnnp.72.4.530): lesion analysis -- depersonalization (parietal Markov blanket disruption) and derealization (Hopfield attractor signal stripped of emotional salience) have distinct neural signatures. D_eff of cognitive representations may be intact in both. pending_design: threshold geometry in (D_eff, stability) space; what combined certainty score constitutes the gate condition? The epistemic-mapping combined certainty score is a candidate. Gated by Q-022 (dissociation must be demonstrated experimentally before designing the combined gate). Registered 2026-03-23. - id: Q-022 title: "Can D_eff (z_self participation ratio, coherence) and Hopfield pattern stability (z_self familiarity) dissociate in the REE architecture?" claim_type: open_question subject: self_maintenance.deff_stability_dissociation polarity: open status: open evidence_quality_note: | narrow_open_question (2026-03-29): three-regime dissociation experiment design specified in notes: (1) Normal training: low D_eff + high stability; (2) Noise perturbation: high D_eff + low stability; (3) Structured novel perturbation: low D_eff + low stability (targeted basis rotation). Experiment to test whether Regime 3 is empirically distinguishable from Regime 2. No experimental evidence yet -- experiment design requires MECH-119 substrate (now provisional). EXQ-084d FAIL 2/4 (2026-03-30): three-regime dissociation (R1=normal, R2=noise sigma=2, R3=novelty). Stability dissociation confirmed: stab_R2=0.063, stab_R3=0.026 vs stab_R1=0.950 (C2/C4 PASS). D_eff dissociation wrong direction: d_eff_R3=22.6 > d_eff_R2=20.8 (C1 FAIL; novelty increases D_eff rather than targeting subspace). D_eff flat across R1/R2 (C3 FAIL). Q-022 verdict: D_eff and stability do dissociate (move independently), answering Q-022 affirmatively, but D_eff ordering between noise/novelty regimes is opposite of prediction. Evidence direction: mixed. Per-claim overrides applied -- see EXQ-084d manifest. depends_on: - MECH-113 - MECH-118 - MECH-119 - SD-005 location: docs/architecture/approach_avoidance_symmetry.md#q-022 source: - evidence/planning/thought_intake_2026-03-23_epistemic_self_monitoring.md - evidence/planning/literature_synthesis_2026-03-23_self_monitoring_dissociation.md notes: > Critical architectural decision question for MECH-113 and the INV-033 implementation. If D_eff and Hopfield stability never dissociate (low D_eff always cooccurs with high stability, high D_eff with low stability), then either metric suffices for MECH-113 self-monitoring -- they are converging measures of the same underlying latent quality. If they dissociate, they are measuring different things and MECH-113 must eventually be split into MECH-118 (familiarity) and MECH-113/D_eff (coherence) as separate mechanisms, each with independent experimental support. Dissociation can be tested by inducing the three regimes (see MECH-119): (1) Normal training: expect low D_eff + high stability (2) Noise perturbation (EXQ-075 protocol): expect high D_eff + low stability (3) Structured novel perturbation: perturb z_self into a coherent but never-before-seen configuration (e.g., targeted basis rotation) -- predict low D_eff + low stability If Regime 3 can be induced and is distinguishable from Regime 2 on both metrics, the dissociation is confirmed. This also validates MECH-119. The Markov blanket framing adds a third axis: can an agent have low D_eff, high stability, but high FEP self-surprise (coherent, familiar, but generating surprising self-states)? This three-way dissociation question is registered as EVB-0069. It determines whether D_eff, Hopfield stability, and Markov blanket resistance are measuring the same latent or three independent dimensions of self-model quality. Literature: The three-way dissociation has been confirmed theoretically and empirically in a dedicated literature synthesis (evidence/planning/literature_synthesis_2026-03-23_ self_monitoring_dissociation.md). Key findings: Friston 2013 (PMID 23825119) -- Markov blanket is topological/structural, defined on different mathematical objects from D_eff (geometric). Parr et al. 2021 (PMID 34573730) -- memory persistence (Hopfield familiarity) is independent of blanket integrity (current conditional independence). Hipolito et al. 2021 (PMID 33607182) -- blanket = connectivity architecture, not representation geometry. Aggleton et al. 2005 (PMID 16154457) -- familiarity survives hippocampal damage (familiarity vs recollection doubly dissociable). Rigotti et al. 2013 (PMID 23685452) -- participation ratio measures capacity, not attractor proximity: D_eff and familiarity are orthogonal by construction. Result: Q-022 shifts from "is dissociation possible?" to "does dissociation occur in z_self during CausalGridWorldV2?" (EVB-0069 experiment design). Added to evidence backlog EVB-0069. Registered 2026-03-23. - id: MECH-116 title: "E1's LSTM hidden state serves as the working memory substrate for goal representation; when conditioned on z_goal_latent, E1 maintains goal context recurrently across steps without requiring ongoing benefit signals." claim_type: mechanism_hypothesis subject: e1.lstm_goal_working_memory polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-089 - MECH-112 location: docs/architecture/approach_avoidance_symmetry.md#mech-116 source: - evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md notes: > E1's LSTM hidden state is the computational analog of DLPFC sustained firing during delay periods and frontal-hippocampal theta coupling during goal-directed navigation. When conditioned on z_goal_latent (MECH-112), E1 incorporates goal context into its recurrent dynamics and maintains this context across steps without the goal needing to be refreshed by ongoing benefit signals -- working memory in the functional sense. Testable prediction: E1 prediction error on goal-approaching trajectory segments is measurably lower when E1 is goal-conditioned vs. unconditioned (E1 predicts better because it knows what it is predicting toward). Non-goal segments are unaffected. Biological grounding: DLPFC neurons sustain elevated firing during delay periods in delayed-response tasks (Fuster & Alexander 1971, Goldman-Rakic 1995). Frontal- hippocampal theta synchronisation (Hyman et al. 2010) is the coupling mechanism: frontal theta signals "navigate toward this goal" to hippocampus. In REE: E1 theta- rate output (via MECH-089 ThetaBuffer) carries the goal context to E3. ARC-032 formalises this communication pathway. Implementation: E1 receives z_goal_latent via a goal_input_proj linear layer that projects [z_self, z_world, z_goal] down to 64-dim before the LSTM -- preserving all existing LSTM dimensions while adding goal conditioning. When e1_goal_conditioned is False, z_goal input is omitted (backward compatible). Note on E1 frontal analogy: the frontal lobe also includes OFC (expected value), ACC (goal PE, conflict monitoring), vmPFC (decision value), premotor/SMA (sequence planning). These additional frontal functions may require further E1 specialisation in V4 beyond goal working memory. Registered 2026-03-23. Experiment: EXQ-076. evidence_quality_note: | EXQ-076d FAIL 1/4 (2026-03-27, 2 runs identical): halflife_ratio=1.0, resource_rate_gap=0, goal_norm_t1200_diff=0. Only C3 PASS (goal_norm > 0 in wanting condition). Root cause: halflife threshold (30% of peak goal_norm) never reached in either condition. At 2000 total steps with 1000-step post-removal window, goal persists robustly in both conditions -- insufficient time to observe meaningful decay. Not a refutation; budget too small. Need 10,000+ steps with goal removal at step 5000 for a meaningful halflife comparison. Queue EXQ-076e with longer training budget. EXQ-076e FAIL 0/3 (2026-03-27, 10000 steps): Extended budget with removal at step 5000. Both conditions IDENTICAL: resource_rate=0.249, halflife=10000, peak_goal_norm=0.3613, goal_norm_at_step6000=0.0014. All metrics equal to 5 decimal places. The E1 conditioning makes no measurable difference to goal maintenance. Two interpretations: (1) z_goal representation is too weak to persist regardless of E1 conditioning (dependent on SD-012 seeding quality, not E1 recurrent maintenance); (2) The benefit signal is insufficient to establish a meaningful goal prior before removal at step 5000. Goal_norm_at_step6000 =0.0014 in BOTH conditions -- goal has already decayed to near-zero. E1 conditioning cannot maintain what wasn't there. Blocked by SD-012 / seeding quality, not testable until z_goal seeding is reliable. EXPERIMENTAL DESIGN PROBLEM (2026-03-29): goal_norm collapses to ~0.0014 in BOTH conditions by step 6000, before the 30%-of-peak halflife threshold is reached. Both conditions report halflife=10000 (never reached), making conditions appear identical. Fix options: (a) longer post-removal window (goal removal at step 7000+ in 15000-step run); (b) lower threshold to 10% of peak; (c) direct exponential decay rate fitting instead of halflife threshold. New EXP proposals queued (2026-03-29) for all three options. EXQ-076e superseded by 076f (2026-04-04): 076e had E1Config.goal_dim never set by from_dims() -> goal_input_proj=None in both conditions -> E1 was goal-agnostic in both. EXQ-076f FAIL 0/3, non_contributory (2026-04-04): bug fixed (goal_dim=32, goal_input_proj active in conditioned arm). Results near-identical: conditioned peak_goal_norm=0.3793 vs unconditioned=0.3806 (0.3% difference); goal_norm_at_r1000=0.0024 vs 0.0025 (both essentially zero). Zero measurable difference in goal persistence. Test design issue (2026-04-04, user decision): the experiment measures z_goal_norm persistence directly. MECH-116 predicts improved E1 prediction quality on goal-directed trajectory segments, not z_goal decay rate. z_goal has a separate decay_goal exponential decay parameter independent of E1 conditioning. Correct test: compare E1 prediction error on goal vs neutral trajectory segments. Redesign required; 076f classified non_contributory (does not test the claim's stated mechanism). - id: MECH-117 title: "The wanting signal (goal-distance scoring via z_goal_latent) and the liking signal (benefit_eval_head receipt predictor) are functionally dissociable in trajectory scoring." claim_type: mechanism_hypothesis subject: wanting.liking_trajectory_dissociation polarity: asserts status: stable implementation_phase: v3 depends_on: - MECH-112 - MECH-116 location: docs/architecture/approach_avoidance_symmetry.md#mech-117 source: - evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md notes: > Wanting (incentive salience, Berridge 1996) and liking (hedonic impact) are dissociable at the neural level: wanting is mesolimbic dopamine (NAc core, VTA), liking is mu-opioid (NAc shell hedonic hotspots). Addicts can want drugs they no longer like. Schizophrenia negative symptoms show intact liking with impaired wanting (Barch & Dowd 2010, Culbreth et al. 2023). In REE trajectory scoring: liking (benefit_eval_head) trains on benefit_exposure at receipt; learns "what does z_world look like when benefit is occurring?"; sharp proximity spike near resource cell; redirects rapidly when resource moves. Wanting (z_goal_latent distance): goal_proximity = 1/(1 + MSE(z_world, z_goal)); continuous spatial gradient rising as agent approaches goal; persists after resource relocation because z_goal decays slowly (half-life ~140 steps at decay_goal=0.005); incentive salience outlasts hedonic change for O(1/decay_goal) steps. The gradient SHAPE distinguishes them: wanting rises smoothly from 5+ steps away; liking spikes sharply only very near receipt. Resource relocation (L1->L2) at step 1500: wanting maintains approach toward L1 for >=25% of next 200 steps; liking redirects to L2 within ~10 steps. Architecture implication: benefit_eval_head is a liking predictor. True wanting requires z_goal_latent. Both signals serve different roles: wanting shapes approach trajectories prospectively; liking calibrates z_goal via benefit_exposure updates. Registered 2026-03-23. Experiment: EXQ-074b. evidence_quality_note: | Same dissociation evidence as MECH-112. EXQ-074d/074e C1 FAIL is a test design confound (greedy nav uniform across conditions). C2 PASS (liking l2_redirect <=7 steps in 074d -- fast redirect to moved resource confirms benefit_eval_head responds to current location, not memory). C3 PASS (wanting l1_fraction=0.355 -- directional approach toward goal location). Wanting and liking operate on distinct spatial and temporal scales as predicted. C1 redesign needed for behavioural resource-rate evidence. EXQ-074e FAIL runs (x3) and EXQ-074f FAIL runs (x3): marked non_contributory 2026-04-06. These runs failed C1 (wanting resource rate gap < threshold) because the serotonergic substrate (MECH-186/187/188) was absent. Without MECH-187 (incentive salience gain regulation), z_goal-driven wanting produces insufficient resource-rate uplift -- the failure is a missing substrate issue, not evidence against MECH-117. These runs do not score toward MECH-117 governance. Re-run after serotonergic mechanisms are implemented (see EXP-0095 to EXP-0101). Clinical application (INV-053): the architecture predicts anhedonia specifically as wanting failure with liking potentially preserved -- the Berridge dissociation is an architectural consequence of z_goal failure (MECH-187 gain suppression), not a general hedonic deficit. EXQ-085g FAIL 3/4 (2026-03-29): wanting-liking dissociation not yet testable because z_goal not pointing to resources (goal_resource_r=0.066). Gated on SD-015 z_resource separation. EXQ-074f PASS 4/4 (2026-04-04): first full behavioral dissociation PASS. wanting l1_fraction=0.96 vs liking l1_fraction=0.43 (delta=0.53 > 0.10 threshold). Liking redirects to L2 in 2 steps; wanting persists toward L1 post-relocation. Consistent with Berridge (1996) incentive salience / Culbreth et al. (2023) profiles. Single seed (42); duplicate run confirms data but not independent replication. EXQ-234 PASS 4/4 (2026-04-04): independent replication with seeds [1,2,3] (EXQ-074f used seeds [42,7,13]). wanting_l1_fraction=1.0 across all 3 seeds (vs liking 0.40-0.66). l1_dissociation=0.34-0.60 (all > 0.10 threshold). goal_norm=0.32-0.37 (seeding active). Combined with EXQ-074f: 6 independent seeds all PASS. No genuine weakens evidence (all prior FAIL/mixed runs reclassified non_contributory -- see above). Dissociation is robust and replicable. Basic mechanism validated. Serotonergic scope: full clinical validation of wanting under adversive conditions (LONG_HORIZON, INV-053) is BLOCKED pending MECH-186/187/188 implementation. EXP-0095 to EXP-0101 define the serotonergic validation experiments. status_note: | Promoted candidate -> provisional 2026-04-04 (EXQ-074f PASS 4/4, seed=42). Promoted provisional -> stable 2026-04-04 (EXQ-234 PASS 4/4, seeds=[1,2,3]). EXQ-234 is an independent 3-seed replication of EXQ-074f (3-seed). Combined: 6 independent seeds all pass 4/4 criteria. wanting_l1_fraction=1.0 all seeds; liking redirects within 2-11 steps of relocation. No genuine weakens evidence (all prior mixed entries had documented design confounds -- greedy nav uniform across conditions). 0 conflict_ratio. conf=0.887. Meets stable threshold. - id: ARC-032 title: "The theta-rate packaging of E1 output (MECH-089 ThetaBuffer) is the primary pathway through which E1's goal-context maintenance reaches E3's trajectory scoring -- the computational analog of frontal-hippocampal theta synchronisation." claim_type: architecture_hypothesis subject: architecture.theta_frontal_hippocampal_goal polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-089 - MECH-116 location: docs/architecture/approach_avoidance_symmetry.md#arc-032 source: - evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md notes: > Frontal-hippocampal theta synchronisation (Hyman et al. 2010, Benchenane et al. 2010) is the biological mechanism by which the frontal goal representation is communicated to hippocampal navigation. PFC theta phase-locks to hippocampal theta during goal-directed navigation; without this coupling, hippocampal navigation becomes undirected. In REE: E1 (frontal working memory, MECH-116) maintains z_goal context in its LSTM hidden state. E1 output is theta-packaged via MECH-089 ThetaBuffer before reaching E3 (hippocampal navigator). The theta channel is therefore the goal communication pathway. This is architecturally appropriate: goal context is a slow signal (maintained across many fast steps), consistent with theta-rate delivery. Testable prediction: if E1 output bypasses the theta buffer (direct fast connection), goal maintenance quality degrades -- shorter persistence, noisier goal-proximity scores. The theta packaging provides smoothing that stabilises z_goal influence. Corollary: ARC-032 is why MECH-116 does not require a separate communication channel for goal context -- the existing theta architecture (MECH-089, SD-006, EXQ-052b) provides the right channel. No new wiring needed beyond E1 goal conditioning. Note: ARC-032 is currently tested jointly with MECH-116 in EXQ-076. A separate theta-bypass ablation would be needed to isolate ARC-032 specifically. Registered 2026-03-23. Experiment: EXQ-076 (joint with MECH-116). evidence_quality_note: | EXQ-076d FAIL 1/4 (tested jointly with MECH-116, 2026-03-27): same null result as MECH-116 at 2000 steps. ARC-032's specific prediction (theta-bypass degrades goal maintenance) has not been tested -- EXQ-076 only tests joint goal conditioning; no theta-bypass ablation condition was included. ARC-032 remains untested as a standalone architectural claim. Design a separate ablation experiment isolating the theta channel (E1 output -> E3 direct vs. E1 output -> ThetaBuffer -> E3). EXQ-076f non_contributory (2026-04-04): EXQ-076 series (076d/e/f) tests E1 goal conditioning via z_goal persistence metrics -- does not isolate the theta-bypass pathway. ARC-032 remains untested as a standalone claim. No evidence_direction change. GOVERNANCE META (2026-04-06): Illusory conflict resolution risk -- zero genuine evidence. EXQ-228, EXQ-228a both non-contributory (z_goal precondition failure: goal_norm=0.031 < 0.05 threshold). Theta pathway has never been testable. Conflict ratio is undefined, not resolved. Any promotion signal would be illusory. Status: pending_retest_after_substrate (gate: EXQ-247 must produce z_goal_norm >= 0.05). - id: ARC-033 title: "SD-003 counterfactual attribution requires a dedicated harm forward model E2_harm_s(z_harm_s, action) -> z_harm_s_next operating within the sensory-discriminative harm stream." claim_type: architectural_commitment subject: harm_stream.sensory_discriminative_forward_model polarity: asserts status: candidate implementation_phase: v3 claim_level: architectural depends_on: [SD-011, ARC-027, SD-010] functional_restatement: > The SD-003 counterfactual pipeline requires z_harm_cf = E2_harm_s(z_harm_s, a_cf): a forward model on the sensory-discriminative harm stream that predicts how harm proximity changes under counterfactual action. This is distinct from and does not require E2.world_forward -- the harm stream is architecturally independent of z_world (SD-010). The forward model is learnable because z_harm_s (proximity/location) has predictable action-conditional structure: moving away from a hazard reduces proximity. With E2_harm_s, causal_sig = E3(z_harm_s_actual) - E3(z_harm_s_cf) operates entirely within the harm stream with no cross-stream bridge. The EXQ-030b validation of SD-003 (world_forward_r2=0.947) confirms the counterfactual pipeline architecture works -- it now needs to be applied to z_harm_s specifically rather than z_world. evidence_quality_note: | EXQ-030b PASS (2026-03-18): SD-003 counterfactual pipeline validated with z_world and E2.world_forward. This confirms the E2-forward-model-as-counterfactual architecture is sound; the same pattern now applies to z_harm_s. EXQ-093 FAIL (bridge_r2=0) and EXQ-094 FAIL (100x regression) confirm that z_world -> z_harm bridge is not the path forward. EXQ-095a FAIL/SLOW_LEARNING (2026-03-27, 900 phase1 eps): Separate FULL (51d) vs HAZARD-ONLY (25d) HarmForwardModel trained. C1 PASS (grad=0.065), C2 PASS (loss 4x drop: 0.00426->0.00113), C5 PASS (causal_approach=0.003864). C3/C4 FAIL: R2=0.0000 for both. hf_hazard_loss shows genuine monotonic decrease throughout 900 eps -- learning IS occurring but model converges to conditional mean without capturing action-conditional variance structure. Not an architectural failure; 900-ep budget insufficient. EXQ-095b with 1800 phase1 eps queued. EXQ-095b FAIL/SLOW_LEARNING_PERSISTENT (2026-03-27, 1800 phase1 eps): R2=0 persists. hf_hazard_loss reaches ~0.0001 (very tiny) but R2=0. Diagnosis: harm_obs_s variance is near-zero across steps (agent mostly not being hit; sparse signal). Model predicts conditional mean with near-zero MSE, yielding R2=0 trivially. More episodes are unlikely to help. Root cause: harm_obs_s (harm_exposure scalar) is too sparse to build action-conditional predictive structure at single-step resolution. Alternative target needed -- possibly harm_obs_a (affective accumulator) after fixing normalization bug (see SD-011 note). EXQ-095c will test forward model on harm_obs_a once SD-011 normalization fix is confirmed by EXQ-101. EXQ-166 MIXED 4/6 (2026-03-29): obs-space forward model (Approach B, residual architecture). Fixes EXQ-115 identity collapse. Seed 42: obs_fwd_r2=0.867 (strong convergence -- architecture validated). Seed 123: obs_fwd_r2=0.0 (P0 encoder quality failure). Causal gap directionally correct (approach > none). Architecture sound; stability issue. EXQ-166a queued with 4 seeds, P0 extended to 80 episodes. EXQ-166a FAIL 4/6 (2026-03-30): 4-seed result. obs_fwd_r2 bimodal (0.925/0.0/0.0/0.885); C2 delta_approach=0.0006 (near zero causal gap). Architectural diagnosis: obs-space approach does not provide selective pressure on harm-causal structure -- MSE loss dominated by position-change prediction, harm dims diluted. E2_harm_s dedicated latent stream (ARC-033) is the correct approach; obs-space is a dead end for C2. EXQ-166b/c/d INCONCLUSIVE (superseded, 2026-03-30): Latent+reconstruction approach. C1 inconsistent (R2=0.028-0.621 across seeds); C2 FAIL in all runs (delta_approach negative -- trained causal signal inverted vs ablated). Reconstruction branch prevents identity collapse in high-R2 seeds but does not recover causal discrimination. Dead end. EXQ-166e PASS 6/6 (2026-03-30): Harm-delta predictor -- predict scalar delta (harm_obs[12]_next - harm_obs[12]_t) from (z_harm_s, action). Identity collapse structurally impossible (delta sign is action-dependent at approach events). delta_r2=0.641 (C1 PASS), causal_gap_approach=0.005 (C2 PASS), causal_gap_contact=0.035 >> causal_gap_neutral=-0.027 (C3 PASS). n_approach_eval=7515 (C4/C5/C6 all PASS). First full PASS for ARC-033 harm forward model. E2_harm_s harm-delta architecture validated. EXQ-195 FAIL (2026-04-04): attribution_gap=-0.044 (C1 FAIL all 4 seeds); harm_forward_r2=0.914 (C4 excellent). Classified non-contributory (2026-04-06 governance): high r2 confirms forward model component working (ARC-033 core validated by EXQ-166e). Negative gap reflects missing SD-003 counterfactual pipeline wiring, not ARC-033 failure. notes: > E2_harm_s is a small MLP: (z_harm_s, action) -> z_harm_s_next, trained with direct MSE supervision on actual next-step z_harm_s values. Architecture mirrors E2.world_forward. z_harm_a (affective stream, SD-011) does NOT need a forward model -- it is not used in counterfactuals. Registered 2026-03-24. See thought intake: evidence/planning/thought_intake_2026-03-24_dual_nociceptive_streams.md - id: INV-034 title: "An agent that only avoids harm but cannot sustain prospective goal-directed motivation cannot exercise genuine agency; goal maintenance is a necessary co-condition for ethical agency alongside harm-avoidance." claim_type: invariant subject: ethical_agency.goal_maintenance_necessary polarity: asserts status: candidate depends_on: - INV-032 - ARC-030 - MECH-112 - Q-021 location: docs/architecture/approach_avoidance_symmetry.md#inv-034 source: - evidence/planning/literature_synthesis_2026-03-22_approach_avoidance_drives.md notes: > INV-032 established that moral agency requires both approach and avoidance drives; pure avoidance produces a degenerate risk manager. INV-034 sharpens this: the required approach drive must be prospective and persistent (wanting, MECH-112/116), not merely reactive hedonic response (liking). An agent with intact liking but absent wanting is stimulus-response, not goal-directed -- precisely the schizophrenia negative symptom profile: intact pleasure when things happen, absent drive to make things happen (Culbreth et al. 2023). Ethical agency in REE requires: (1) harm-avoidance (SD-010, MECH-095, ARC-016); (2) goal-directed approach (MECH-112, MECH-116, ARC-030). Both are necessary because harm-avoidance alone produces quiescence (Q-021) -- the minimum-action policy that causes no harm because it does nothing. Quiescence is not ethical agency; it is paralysis. The D1/D2 commit threshold (ARC-030) = balance between approach value (D1/wanting) and harm cost (D2). An agent with no D1 signal cannot commit through positive drive; it can only commit by default when harm is absent. This is not decision-making. Note: INV-034 does not claim liking is unnecessary -- liking (benefit_eval_head) is the learning signal that calibrates z_goal (wanting). The claim is about what is necessary for genuine agentive commitment, not about computational signals involved. Clinical grounding: avolition in schizophrenia, anhedonic depression, amotivation syndromes all demonstrate that harm-avoidance systems remain intact while agency collapses. REE must avoid this architecture failure mode. Registered 2026-03-23. evidence_quality_note: | INV-034 depends on MECH-112/116 (goal maintenance architecture). EXQ-085 through EXQ-085e all failed at the z_goal seeding prerequisite (SD-012 homeostatic drive bottleneck). INV-034's core claim (goal maintenance is necessary for ethical agency) is conceptually uncontested but experimentally untestable until SD-012 is resolved. Path forward: same as MECH-071/SD-012. Hold pending architectural discussion of z_goal seeding mechanism. Pathological inverse (INV-053): failure of goal maintenance under adversive conditions is the computational definition of the depressive attractor state (EXQ-237a). The claim now has both a normative face (goal maintenance necessary for ethical agency) and a clinical face (goal maintenance failure = depression). - id: INV-035 title: "A REE state must not be defined purely by sensory appearance; two perceptually identical situations that differ in temporal position, active commitment, goal/antigoal relation, social context, or operative constraints constitute distinct states." claim_type: invariant subject: state.not_raw_perception polarity: asserts status: candidate depends_on: - ARC-004 - ARC-003 - ARC-007 location: docs/architecture/state.md source: - evidence/planning/thought_intake_2026-03-24_state_definition_hippocampal_primitives.md notes: > Grounded in the formal state definition ("situation as navigable from here"). A state definition that collapses to perceptual embedding silently breaks trajectory evaluation, viability mapping, and ethical reasoning. The same sensory input at a door produces entirely different viable transitions, costs, antigoals, urgency, and commitment structures depending on temporal position (T), active constraints (C), and goal/antigoal relations (G, A). Collapsing these to appearance means the hippocampal rollout cannot distinguish "leaving for work" from "fleeing danger" -- two trajectories with nearly opposite valid next-action sets. This invariant is a design guard against the common shortcut of treating latent perceptual embeddings as sufficient state representations. Registered 2026-03-24. - id: INV-036 title: "A REE state is valid only if it supports transition prediction, valence and antigoal tagging, and uncertainty representation; a representation lacking any of these cannot function as a navigable state." claim_type: invariant subject: state.functional_requirements polarity: asserts status: candidate depends_on: - INV-035 - ARC-007 - ARC-018 - ARC-003 location: docs/architecture/state.md source: - evidence/planning/thought_intake_2026-03-24_state_definition_hippocampal_primitives.md notes: > The three functional requirements are the minimum for a representation to participate in hippocampal rollout and E3 path evaluation. (1) Transition prediction: without plausible-next-transition grounding (R component), the hippocampal system cannot chain the state into a rollout. (2) Valence and antigoal tagging (G, A components): without goal-relation and antigoal-relation fields, E3 has no directional signal for path selection -- trajectories cannot be ranked as better or worse relative to any objective. (3) Uncertainty (U component): without structured uncertainty, precision weighting in the control plane has nothing to operate on, and epistemic caution cannot propagate into trajectory evaluation. These three requirements are necessary but not sufficient -- a complete state also requires W, Self, T, and C (world, self, temporal position, constraints). INV-036 specifies the floor below which a representation is definitely not a state; INV-035 specifies that perceptual identity is insufficient for state identity. Together they define the outer boundary of valid state abstraction. Registered 2026-03-24. - id: INV-037 title: "A content class that is stored and retrievable in the REE system is not thereby active in the navigable state used for trajectory evaluation; active participation requires a preparation substrate (vmPFC-analog) that converts stored content into live state components at evaluation time." claim_type: invariant subject: state.stored_vs_active_distinction polarity: asserts status: candidate depends_on: - INV-035 - INV-036 - ARC-035 - ARC-003 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > The stored/active distinction is the central architectural fact that ARC-035 explains. Information can be stored in memory, retrievable on demand, expressible in propositional form -- and still not participate in trajectory evaluation because the preparation substrate that would make it a live state component is absent or dysfunctional. This is not equivalent to INV-035 (state not defined by sensory appearance) or INV-036 (state functional requirements). Those invariants specify what state must contain and how it is bounded. This invariant specifies a procedural requirement: content does not become a state component automatically by virtue of being stored -- a dedicated preparation step is required. Primary evidence: the knowledge/application dissociation. Patient EVR (Eslinger/Damasio 1985, PMID 4069365) had above-average IQ, intact language, intact declarative memory, and could correctly describe appropriate choices in social scenarios -- while continuously making catastrophically inappropriate choices. The correct content was stored and retrievable; it was not active in the navigable state used for path selection. vmPFC ablation had removed the preparation substrate. Experimental predictions: (1) A system with correct residue storage but degraded preparation routing will show intact residue retrieval in explicit queries while failing to avoid harm-associated trajectories in generation (stored but not active). (2) A system with correct constraint encoding but no preparation pathway will articulate constraints correctly when queried while generating constraint-violating trajectories. (3) The failure mode is not correctable by adding post-hoc constraint checking -- see INV-038. Registered 2026-03-25. - id: INV-038 title: "A system with correct post-hoc ethical scoring but without an active constraint preparation substrate will produce the EVR pattern: correct verbal moral judgments coexisting with unconstrained trajectory generation; this pattern is not correctable by improving post-hoc scoring accuracy." claim_type: invariant subject: ethics.post_hoc_filter_insufficiency polarity: asserts status: candidate depends_on: - INV-001 - INV-037 - ARC-035 - ARC-003 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > INV-001 asserts that ethical behavior cannot be compiled into a single explicit ethics module. INV-038 is more specific: it asserts that a post-hoc scoring filter -- applied to trajectories after they are generated -- cannot produce ethical behavior when the generative state lacks active constraints, regardless of how accurate the scoring is. The argument: post-hoc filtering operates on trajectories generated by the state. If the state lacks active constraints (because the preparation substrate is absent), the generated trajectories encode unconstrained proposals. A filter that correctly identifies and suppresses these proposals does not change the fact that (a) unconstrained trajectories were generated as the primary candidates, (b) suppression under time/resource pressure fails first, (c) partial or ambiguous cases default to the unconstrained trajectory, (d) the system's behavior under novel constraint configurations is unpredictable because the constraint was never in the generative state. EVR grounding: the patient's behavior was continuously norm-violating despite intact verbal moral reasoning (a post-hoc system). Adding more accurate verbal scoring would not have changed his trajectory generation -- it would have changed his ability to explain why his choices were wrong, which is not the same thing. Experimental falsifiability: a system with vmPFC-equivalent preparation intact but degraded post-hoc scoring should show ethical behavior with poor explicit moral reasoning. A system with correct post-hoc scoring but degraded preparation should show the EVR pattern. If only the first is observed and not the second, INV-038 is weakened. This invariant has direct implications for AI alignment: inserting an ethical scoring module downstream of trajectory generation does not solve the problem that INV-038 identifies. The solution requires that normative content be active in the generative state, not evaluated after generation. Registered 2026-03-25. - id: MECH-120 title: "The SWS sub-phase of offline mode performs global attractor flattening and residue field denoising (synaptic homeostasis analog), pruning weak noisy traces and restoring signal-to-noise ratio." claim_type: mechanism_hypothesis subject: sleep.sws_denoising_attractor_flattening polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-030 - MECH-092 - MECH-094 - ARC-016 location: docs/architecture/sleep/offline_phases.md#mech-120 source: - evidence/literature/targeted_review_connectome_mech_030/entries/tononi_cirelli_2006/summary.md notes: > Biological basis: Synaptic Homeostasis Hypothesis (Tononi & Cirelli 2006). Slow wave sleep globally downscales synaptic weights toward a principled baseline, reversing unsupervised Hebbian potentiation from waking. Effect: weak/noisy traces pruned, strong traces sharpened, SNR restored. SWS is the earliest offline sub-phase. REE analog: during the SWS sub-phase of MECH-030, three operations occur: (1) Residue field global decay/normalisation -- residue magnitudes scaled toward baseline, reducing attractor depth for all trajectory biases (prevents any single harm or goal trace from accumulating indefinitely). (2) z_delta recalibration -- the deepest latent (z_delta) is pushed toward a principled low-information prior, resetting the long-horizon world model before NREM replay repopulates it with consolidated content. (3) Low-confidence representation pruning -- traces below a salience threshold are zeroed, freeing representational capacity for consolidation. Critical constraint (Walker 2004): SWS downscaling is non-selective by default. Without MECH-094 hypothesis tag, genuine harm signals are indistinguishable from noise and will be downscaled along with it. MECH-094 is a HARD PREREQUISITE for MECH-120: the tag marks which traces must be protected from SWS normalisation. V3 static setpoints that V4 will make dynamic (see v3_v4_transition_boundary.md): residue field decay rate (currently static per-step), z_delta recalibration magnitude (currently none), SNR pruning threshold (currently absent). V3 must measure these setpoints under normal operation to calibrate V4 dynamics. Functional outcome (added 2026-04-04, grounded in Hagewoud 2010 + Suppermpool 2024): The primary functional output of MECH-120 is behavioral strategy diversity preservation, not merely SNR restoration. Without SHY-equivalent downscaling, Hebbian winner-take-all dynamics during waking produce monostrategy convergence even in benign environments (no harm conditioning, no MECH-124 pathway) -- the dominant action trajectory monopolises the synapse budget, and the flexible map-based alternative becomes inaccessible. This is the neutral-environment Hebbian lock-in variant, distinct from MECH-124 (harm-trace-driven contraction). Empirical grounding: Hagewoud et al. 2010 (Sleep) showed that sleep deprivation shifts learning from hippocampal spatial strategy to rigid striatal response strategy in mice, with reversal learning deficit confirming reduced flexibility. evidence_quality_note: > EXQ-245 FAIL x2 (2026-04-05, 2026-04-07): both reclassified to non_contributory (governance 2026-04-08). MECH-120 is V4-scope; V3 substrate lacks MECH-030, MECH-092, MECH-094, ARC-016 prerequisites. s1 (SHY-then-SWS beats SWS-alone) failure reflects absent substrate, not claim falsification. Run 1 showed s2 PASS (correct order beats wrong order, 3/3 seeds) but run 2 regression (s2 also FAIL 1/3). Literature support strong (7 entries). Held at candidate pending V4 substrate implementation. Note (2026-04-08): user confirmed decision to pull V4 mechanisms into V3 scope -- MECH-120 substrate (SHY decay, SWS residue scaling) should be planned for V3 implementation. hold_candidate_resolve_conflict applied. Suppermpool et al. 2024 (Nature) confirmed single-neuron SHY operation in zebrafish with sleep-pressure gating. Prediction for REE: a minimal mind without a sleep substrate will show monostrategy convergence under extended training. This is the currently observed signature in V3 (no sleep substrate) and is predicted by MECH-120 absence, not by any other mechanism. Registered 2026-03-23. V4 scope. Functional outcome note added 2026-04-04. Implementation note (2026-04-08): SHY wiring completed -- enter_sws_mode() now calls E1.shy_normalise(decay) when shy_enabled=True. Config: REEConfig.shy_enabled (bool, default False), REEConfig.shy_decay_rate (float, default 0.85). EXQ-245a queued for validation. - id: MECH-121 title: "The NREM sub-phase of offline mode replays recent episodic content via SWR-equivalent sequences, transferring episodic traces from E3/hippocampal buffer to E1/neocortical schematic representation." claim_type: mechanism_hypothesis subject: sleep.nrem_swr_replay_consolidation polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-030 - MECH-092 - MECH-094 - MECH-122 - SD-004 location: docs/architecture/sleep/offline_phases.md#mech-121 source: - evidence/literature/targeted_review_connectome_mech_030/entries/diekelmann_born_2010/summary.md notes: > Biological basis: Diekelmann & Born 2010. During NREM, hippocampal sharp-wave ripples (SWRs) replay recent episodic content in compressed form (~20x compression). SWR replay is coordinated with thalamo-cortical spindles (MECH-122) and cortical slow oscillations (<1 Hz) in a three-way dialogue: slow-oscillation up-state triggers spindle, spindle gates SWR, SWR content travels to neocortex for integration into long-term schematic knowledge. REE analog: MECH-092 (quiescent E3 heartbeat replay, V3) is the functional prerequisite. MECH-121 extends MECH-092 into the full consolidation pipeline: E3/HippocampalModule SWR-equivalent replay (already started in MECH-092) is coordinated with z_theta->z_delta delivery (spindle analog, MECH-122), and E1's LSTM (neocortical analog) integrates the replayed content into its long-horizon schematic world model. The consolidation transfers specific goal-directed episode traces (approach sequences, harm encounters) from E3's episodic buffer to E1's generalised world model. Two important properties of this transfer: (1) Temporal compression: NREM replay is ~20x faster than real-time. In REE terms, E3's SWR replay runs at fast rollout speed without the real-time constraint. This means the MECH-121 phase can consolidate many episodes in a single offline cycle. (2) Semantic abstraction: hippocampus holds episodic traces; neocortex extracts schematic regularities. The consolidation is NOT lossless replay -- it is abstraction. E1 should integrate the invariant patterns across replayed episodes, not store each episode verbatim. This is why E1's goal conditioning (MECH-116) improves with sleep: goal-relevant patterns get abstracted into E1's world model. V3 prerequisite: MECH-092 quiescent replay validated. Balanced replay scheduling (harm AND goal traces, not harm-dominated) must be implemented before MECH-121 to prevent MECH-124 consolidation-mediated contraction. Registered 2026-03-23. V4 scope. - id: MECH-122 title: "Thalamo-cortical spindle-equivalent bursts package E3/hippocampal replay content for z_theta delivery AND gate external sensory input, protecting the consolidation episode from interruption." claim_type: mechanism_hypothesis subject: sleep.spindles_packaging_gating polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-030 - MECH-089 - MECH-121 - SD-006 location: docs/architecture/sleep/offline_phases.md#mech-122 source: - evidence/literature/targeted_review_connectome_mech_030/entries/diekelmann_born_2010/summary.md notes: > Biological basis: Sleep spindles (11-15 Hz thalamo-cortical bursts, 0.5-3 sec duration) during NREM stage 2. Spindles serve two distinct functions: (1) Content packaging: spindles gate hippocampal SWR replay content to neocortex. The spindle burst is the transmission window -- during the spindle up-phase, hippocampal content is packaged into thalamic relay neurons and forwarded to cortex. Without spindles, SWR replay content does not reliably reach neocortex. (2) Sensory gating: spindles suppress thalamocortical relay of external sensory input during the consolidation window. This makes the offline phase robustly offline -- external stimuli cannot interrupt or contaminate the consolidation episode. REE analog: Content packaging: the ThetaBuffer (MECH-089, EXQ-052b PASS) already provides theta-rate packaging of E1 output to E3. MECH-122 extends this: during offline mode, a spindle-equivalent burst mechanism packages E3/hippocampal SWR replay output for z_theta delivery to E1 (the reverse direction of the waking theta channel). The theta channel is bidirectional: waking = E1->E3 (goal context); sleep = E3->E1 (episodic consolidation). Sensory gating: during MECH-122 spindle windows, the sensory input pathway to z_self and z_world encoders is gated closed. New environmental observations are buffered but not processed. This prevents waking-state observations from contaminating the consolidation episode. V3 implication: the ThetaBuffer (already implemented) is the V3 scaffolding for MECH-122. V3 must ensure the ThetaBuffer can run bidirectionally (E1->E3 waking, E3->E1 sleeping). Currently only waking direction implemented. Note: spindle density correlates with overnight memory consolidation performance in humans (Walker 2004). In REE, spindle density is a proxy for consolidation efficiency -- more spindle bursts per offline cycle = more episodic content transferred to E1. This provides a potential diagnostic metric in V4. Registered 2026-03-23. V4 scope. evidence_quality_note: > EXQ-246 PASS x2 (2026-04-05): two V3 proxy spindle coordination runs. Both reclassified to inconclusive (governance 2026-04-07 for first run, 2026-04-08 for second indexed run). PASSes are vacuously true -- WITH_SPINDLE and NO_SPINDLE conditions produced identical metrics in all 3 seeds in both runs (harm rates, harm discrimination, harm_eval values all identical). Phase 3 spindle proxy (ThetaBuffer consolidation_summary written back to ContextMemory) had zero measurable effect. V3 proxy too weak to produce signal for this V4-scope mechanism. Promotion to provisional deferred: no real experimental evidence. MECH-122 requires a functional spindle mechanism (V4-scope). hold_candidate applied. - id: MECH-123 title: "The REM sub-phase of offline mode recalibrates precision priors and tests world model hypotheses via unconstrained generative simulation, with commit gate fully suppressed." claim_type: mechanism_hypothesis subject: sleep.rem_precision_recalibration polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-030 - MECH-092 - MECH-094 - ARC-016 - MECH-093 location: docs/architecture/sleep/offline_phases.md#mech-123 source: - evidence/literature/targeted_review_connectome_mech_030/entries/hobson_friston_2012/summary.md notes: > Biological basis: Hobson & Friston 2012. REM = unconstrained hyperparameter exploration under aminergic suppression. The predictive coding framing: during REM, precision parameters (priors) are recalibrated without being anchored to current sensory evidence. Hallucinatory content = the generative model running unconstrained, testing hypotheses against internally-generated data. Aminergic suppression (NE, 5-HT near zero) = commit gate fully collapsed. Cholinergic dominance = high internal noise, enabling exploration of prior space. REE analog: the REM sub-phase is the precision recalibration phase. (1) z_beta recalibration: z_beta (NE analog, MECH-059) is driven toward a principled low-arousal prior. The precision weights that z_beta modulates (E3 prediction error gains) are reset toward a non-biased baseline before waking resumes. (2) z_delta hyperparameter reset: the deepest latent (z_delta, encoding long-horizon priors) is recalibrated via E1 running without sensory anchoring. E1's LSTM explores its prior distribution unconstrained by current observation. (3) Commit gate fully suppressed: ARC-016 running_variance threshold is effectively infinite during REM -- no commitment can occur. All E3 activity is hypothesis-tagged (MECH-094) and does not accumulate in the residue field. (4) World model hypothesis testing: E1 generates unconstrained world-model trajectories (the REE equivalent of dream content), testing whether the current world model produces coherent self-consistent trajectories. Incoherent predictions reveal model limitations to be addressed on waking. V3 static setpoints to be measured (for V4 dynamic calibration): commitment_threshold baseline (0.40 -- what does normal waking operation look like?), z_beta EMA alpha (0.3 -- what is the natural arousal time constant?), E1 prediction_horizon (20 -- what horizon does E1 naturally explore in REM?). Critical constraint: MECH-094 hypothesis tag is essential. REM-generated trajectories must be tagged as hypothesis, not real-action. If the tag fails during REM, hallucinatory content accumulates in the residue field as if it were real experience -- the psychosis mechanism (MECH-094 notes). Registered 2026-03-23. V4 scope. - id: MECH-124 title: "When harm-trace salience dominates offline replay content, consolidation selectively amplifies harm predictions, progressively contracting option space -- a distinct failure mode from Q-021 behavioral flatness." claim_type: mechanism_hypothesis subject: sleep.consolidation_mediated_option_contraction polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-030 - MECH-120 - MECH-121 - MECH-092 - MECH-094 - Q-021 location: docs/architecture/sleep/offline_phases.md#mech-124 source: - evidence/literature/targeted_review_connectome_mech_030/entries/walker_stickgold_2004/summary.md notes: > Biological basis: Walker & Stickgold 2004. Sleep consolidation is not neutral -- it amplifies whatever dominates replay content. Fear-conditioned traces are selectively strengthened during REM. In PTSD, REM intrusions re-consolidate the fear trace repeatedly: nightmare replay -> re-consolidation -> stronger fear -> more REM intrusion. Extinction fails because consolidation of the fear trace outpaces extinction of it. The consolidation loop is self-amplifying. REE failure mode (distinct from Q-021): Q-021: behavioral flatness from pure-avoidance TRAINING (no approach signal from the start -- gradient minimum is quiescence). MECH-124: behavioral flatness from consolidation DYNAMICS despite approach signals being present in training. Mechanism: (1) Harm salience > goal salience in residue field -> SWR replay (MECH-121) is dominated by harm traces. (2) Consolidation selectively strengthens dominant replay content -> harm predictions in E1 and E3 amplified. (3) SWS denoising (MECH-120) cannot distinguish goal traces from noise if they are low-salience -> goal representations downscaled. (4) Weakened goal drive (z_goal decay not counteracted by consolidation) -> approach behavior declines. (5) Fewer positive experiences -> even less goal representation -> self-amplifying loop. The loop is broken only by: balanced replay scheduling (MECH-121 must include goal-directed traces proportionally), MECH-094 tag protecting harm traces from over-amplification, and the go mechanism (ARC-030, MECH-112/116/117) providing sufficient approach signal to counteract harm dominance. Clinical analog: PTSD (trauma loop), treatment-resistant depression (learned helplessness + sleep-mediated entrenchment), avolition in schizophrenia where consolidation entrenchment amplifies negative predictions. Distinct from Q-021 but related: Q-021 is the training-time failure (static); MECH-124 is the consolidation-time failure (dynamic, requires sleep architecture). An agent with sufficient approach drives (INV-034) can still fall into MECH-124 if its consolidation system is imbalanced. Prevention requires balanced replay content scheduling as an explicit architectural constraint on MECH-121. V3 implication: even before V4 sleep implementation, V3 must ensure that the residue field and goal representation (z_goal) are balanced in salience. If z_goal salience is systematically lower than harm salience, a V4 consolidation system will amplify this imbalance. EXQ-074b and EXQ-076 provide early evidence on whether z_goal can achieve competitive salience with harm traces. Registered 2026-03-23. V4 scope. evidence_quality_note: | EXQ-224 DIAGNOSTIC (2026-04-04): z_goal salience vs harm salience probe. BASELINE condition (z_goal_enabled=True, drive_weight=2.0), 3 seeds x 300 episodes. Mean final ratio (z_goal_norm / harm_salience) = 0.312 (goal ~31% of harm salience, harm dominates ~3:1). FLAG_2_ratio_declining=True (mean ratio slope=-0.0034; ratio declining over training in 2/3 seeds). risk_detected=True. V4 risk condition confirmed: in the current V3 substrate, harm salience systematically exceeds goal salience; if offline consolidation is harm-dominated in V4, the MECH-124 loop predicts progressive option contraction. Diagnostic only -- does not test the consolidation mechanism itself (no sleep in V3). Evidence_direction not changed; governance decision remains hold_candidate_resolve_conflict (requires V4 sleep architecture experiment). - id: MECH-125 title: "E3 trajectory selection implements multi-constraint viability evaluation rather than scalar reward maximisation: a trajectory is selected when no major evaluation system produces a veto-level error signal across goal space, harm space, identity space, and resource space." claim_type: mechanism_hypothesis subject: coherence.multiconstraint_viability polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-112 - MECH-113 - ARC-018 - ARC-003 - SD-005 location: docs/architecture/e3.md#4a-selection-criterion-multi-constraint-viability-not-reward-maximisation source: - docs/thoughts/2026-03-24_COHERENCE_MULTICONSTRAINT_HIPPOCAMPAL_NAVIGATION_CONVERGENCE.md notes: > The selection rule is: lowest-veto + acceptable goal progress, not maximum reward. A trajectory is coherent when it does not produce a veto-level error signal in any major evaluation system simultaneously. This is distinct from reward maximisation: a trajectory can have high immediate reward while failing harm, identity, or resource constraints, producing internal contradiction. Biological correlate: the commit gate (basal ganglia) opens when ACC conflict, insula cost, amygdala threat, goal-distance (MECH-112), and identity checks all fall below veto threshold simultaneously. No single system "chooses" -- commitment emerges from absence of sustained veto. Distinct from MECH-113 (which measures internal latent coherence via D_eff, a single-space metric). MECH-125 is the cross-space selection criterion: the multi- constraint viability check across goal, harm, identity, and resource spaces that determines whether E3 commits to a trajectory at all. The five terms in E3's selection functional J(tau) map to constraint spaces: C(tau) = cross-horizon coherence (identity + prediction consistency), R(tau) = residue/ethical curvature (harm + moral cost), F(tau) = feasibility (resource + physical constraint), I(tau) = epistemic value (exploration benefit, negative constraint), P(tau) = phase-lock (temporal coherence). The ACC + insula complex is the likely biological substrate for the coherence error signal (conflict monitoring + interoceptive cost). Commitment emerges when these signals collectively fall below threshold -- not from a single coherence module. Convergent neuroscience: Behrens et al. (2018) Neuron "What is a cognitive map?" establishes hippocampal-entorhinal system as a general relational geometry engine; Whittington et al. (2020) Cell Tolman-Eichenbaum Machine unifies spatial and relational cognition. REE arrived at the same architecture independently from the design side. See MECH-112 (goal latent) and MECH-113 (D_eff self-coherence) for the specific sub-mechanisms this claim coordinates. Registered 2026-03-24. V3 scope. - id: MECH-126 title: "Specific modes of REE state abstraction failure -- overmerge, oversplit, temporal context loss, valence mis-tagging, uncertainty collapse, and narrow representational capacity -- produce behavioral signatures that are structurally analogous to empirically described psychiatric conditions, sharing identifiable circuit-level mechanisms." claim_type: mechanism_hypothesis subject: state_abstraction.failure_modes_psychiatric_analogs polarity: asserts status: candidate depends_on: - INV-035 - INV-036 - ARC-007 - ARC-018 location: docs/architecture/state.md source: - evidence/planning/thought_intake_2026-03-24_state_definition_hippocampal_primitives.md - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > Six failure modes identified and grounded in the literature (2026-03-24 pull): (1) Overmerge -> schizophrenia/thought disorder: strongly supported. DG pattern separation failure + CA3 pattern completion excess produces exactly the predicted behavioral signature -- contextually inappropriate responses, intrusion of irrelevant context, source monitoring failure. Evidence: PMID 35853896 (shallow cognitive map hypothesis), PMID 39567329 (Trends Cog Sci 2025 attractor instability), PMID 20810471 (Tamminga hippocampal formation SZ), PMID 17020747 (fronto-hippocampal temporal context monitoring). (2) Valence spreading / attractor basins -> anxiety/PTSD: strongly supported. Hippocampal pattern completion extends threat-tagged zone (GAD). Safety attractor cannot be entered from extinction context in PTSD (deep basin). Evidence: PMID 27794690 (Kaczkurkin PTSD fear generalization fMRI), PMID 31206738 (Lis 2020 generalization review), PMC 7554263 (contextual reinstatement MVPA PTSD), PMC 10728304 (vmPFC-hippocampus-amygdala circuit review Frontiers 2023), Lissek et al. Neurosci Biobehav Rev 2021 meta-analysis. (3) Temporal/contextual component loss -> ADHD: well supported. PFC-hippocampal-cerebellar circuits maintain T (temporal position) and C (constraints); dysfunction degrades both, producing appearance-driven impulsive action. Evidence: PMC 11325328 (White/Dalley 2024 temporal processing ADHD), PMC 2894421 (Arnsten 2009 PFC in ADHD), PMID 19621976 (Arnsten 2009 CNS Drugs), PMID 15071717 and 12424557 (Barkley temporal processing), PMC 3329889 (multi-pathway ADHD model). (4) Uncertainty collapse -> mania: theoretically well-grounded, direct empirical evidence developing. Precision dysregulation in active inference (excessive prior precision = U->0) produces overcommitted, over-confident, reduced-exploration action. Evidence: PMID 32860285 (Smith/Badcock/Friston 2021 predictive coding clinical), PMID 39828236 (Qela 2025 transdiagnostic PC review), PMC 10196365 (computational belief updating in psychotic-like experiences), PMID 29400358 (Huys 2018 computational parameters anhedonia/mood disorders). (5) Narrow representational capacity / negative G tagging -> depression: multiply supported. Reduced state space capacity (synapse loss paper PMID 39569353) is sufficient to produce depression-like behavior without parameter manipulation. Negative learning-rate asymmetry (PMC 5828520 anhedonia review) and inflated path cost estimation (effort learning) are additional mechanisms. Learned helplessness encodes approach as structurally unavailable (PMID 27337390, 26555633). (6) Oversplit -> ASD generalization failure: supported behaviorally, mechanistic account developing. Increased hippocampal pattern separation produces orthogonal representations for structurally similar contexts. Evidence: PMC 7907419 (Pattern Unifies Autism framework), PMC 4573235 (de Marchena 2015 generalization weakness ASD), PMID 27119213 (stimulus overselectivity ASD/DS), DOI 10.1177/1088357620943504 (Kelly/Reed overselectivity ASD). Cross-track convergence: hippocampus is central across all six tracks but in different directions on the pattern separation/completion axis. This axis (DG-CA3 balance) emerges as the primary organizing parameter: too much completion -> overmerge (SZ) + attractor lock (PTSD); too much separation -> oversplit (ASD). ADHD and depression involve orthogonal dimensions (temporal binding, representational capacity/valence). The claim is structural analogy, not identity: REE state components (W, Self, T, G, A, C, U, R) are not identical to empirical constructs, and the mapping does not resolve causal direction. Strongest causal evidence: PMID 39569353 (synapse loss sufficient for depression- like behavior in RL model). (7) Residue non-propagation -> psychopathy: strongly supported. Three-node failure cascade: amygdala generation failure (PMID 15997022 Birbaumer 2005 fMRI; PMID 11704074 Kiehl 2001; PMID 18281412 Marsh 2008 CU-trait children); encoding failure under goal pursuit (DOI 10.1177/0956797610396227 Baskin-Sommers 2011); vmPFC/OFC update failure (PMID 16866595 Budhani 2006 reversal; PMID 26359751 Gregory 2015 Lancet Psychiatry fMRI; PMID 18458210 Finger 2008; PMID 15134853 Blair 2004 OFC). Glenn/Raine 2011 (PMID 21919563) articulates generation vs. propagation as two separable failure nodes. RL mechanism: wide learning window (Ly 2021, PMID 38773992). (8) Social/identity constraint omission -> frontal damage / bvFTD: strongly supported. Knowledge/application dissociation: Saver/Damasio 1991 (PMID 1791934); EVR Eslinger/Damasio 1985 (PMID 4069365). Early-onset variant: Anderson 1999 (PMID 10526345). Constraint removal: Koenigs 2007 (PMID 17377536 Nature). Progressive erosion: bvFTD Rascovsky 2011 (PMID 21810890); Rahman 1999 IGT FTD (PMID 10430832). Affective ToM substrate: Shamay-Tsoory 2003 (PMID 12729486); Moll/Grafman 2005 NatRevNeurosci (PMID 16276356). Three organizing axes span all nine failure modes: (1) hippocampal pattern sep/completion balance; (2) PFC-hippocampal temporal/constraint binding; (3) amygdala-vmPFC valence generation-propagation cascade. vmPFC is the critical integration node across axes 2 and 3. Registered 2026-03-24. Literature pull completed 2026-03-24. - id: MECH-127 title: "Counterfactual other-cost-aversion activates cooperative behavior as a motivational surrogate when the direct task-reward pathway is degraded." claim_type: mechanism_hypothesis subject: social.counterfactual_empathic_activation polarity: asserts status: candidate depends_on: - MECH-031 - MECH-032 - INV-005 - INV-028 - INV-034 location: docs/architecture/social.md#mech-127 source: - docs/thoughts/2026-03-24_empathy_multiagent_ethics.md - evidence/planning/thought_intake_2026-03-24_empathy_multiagent_ethics.md notes: > When an agent's direct task-motivation pathway is degraded (deficit state), modeling the anticipated cost to another agent -- other-cost-aversion -- can substitute as an activation mechanism, producing cooperative task-engagement even in the other agent's absence. The activation is counterfactual: it fires on a model of what would happen to the other agent if they took on the task, not on their actual current state. The resulting distress-at-their- struggle provides motivational energy that the depleted direct pathway cannot supply. Behavioral output (task engagement / cooperation) is identical to the direct-pathway output but arrives via a different activation route. Key distinction from MECH-031: MECH-031 describes empathy modulating behavior via OTHER_SELFLIKE tagging and control-plane coupling -- modulation of an existing pathway. MECH-127 describes empathy replacing a degraded pathway -- bypass, not modulation. Implication for MECH-032 (high-recall OTHER_SELFLIKE bias): counterfactual projection requires OTHER_SELFLIKE tagging to function pre-encounter, supporting the bias rationale. Motivating observation: first-person report, 2026-03-24. Registered 2026-03-24. - id: ARC-034 title: "Ethics testing must span nth-order multiagent trajectory distributions; local pairwise probes are insufficient to characterize emergent ethical properties." claim_type: architectural_commitment subject: ethics.test_scope_nth_order polarity: asserts status: candidate depends_on: - INV-001 - INV-015 - INV-028 - MECH-127 - MECH-129 location: docs/invariants.md#arc-034 source: - docs/thoughts/2026-03-24_empathy_multiagent_ethics.md - evidence/planning/thought_intake_2026-03-24_empathy_multiagent_ethics.md notes: > A REE system can be locally ethical at every pairwise (n=1) interaction and still produce an ethically problematic emergent collective state at n=k. Conversely, locally depleted agents can produce more ethical emergent behavior at n=k than n=1 predicts (as in the MECH-127 motivating case: direct-pathway analysis predicts reduced helping; nth-order analysis predicts increased helping via counterfactual empathic activation). Ethics tests must therefore include: (1) counterfactual probes -- what activates when the direct pathway is blocked? (2) multi-agent stress tests -- does ethical behavior degrade or amplify under load? (3) cascade mapping -- trace the nth-order activation chain, not just the first-order output. Formally: given agents {A_1...A_n}, local rules theta, threshold parameters {x...x_n}, and feedback topology G -- characterize convergence to emergent state q and its ethical properties. This requires descriptive (attractor theory), prescriptive (Lyapunov / potential game), and diagnostic (transfer entropy / perturbation) test types. Does not contradict INV-001 (no explicit ethics module) -- this is a test scope claim, not an architectural claim about where ethics lives. Registered 2026-03-24. - id: ARC-035 title: "vmPFC is the substrate that converts stored affective, normative, and residue content -- prior outcome history, social and identity constraints, learned value associations, and safety memories -- into active components of the navigable state used by E3 during trajectory evaluation." claim_type: architectural_commitment subject: vmPFC.affective_normative_state_preparation polarity: asserts status: candidate depends_on: - ARC-007 - ARC-013 - ARC-003 - INV-035 - INV-036 - MECH-126 location: docs/architecture/vmPFC.md#arc-035 source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md - evidence/planning/thought_intake_2026-03-24_state_definition_hippocampal_primitives.md notes: > The core claim is a functional dissociation: information can be stored in the system and declarable in verbal form while remaining inactive as a component of the navigable state used for trajectory evaluation. vmPFC is the substrate that bridges this gap for a specific content class: affective residue from prior outcomes, social and identity constraints, learned safety memories, and goal approach pull. Four failure modes grounded in the MECH-126 literature pull: (1) Residue non-propagation (psychopathy): prior punishment history exists but does not update future trajectory selection -- reversal learning deficit (PMID 16866595 Budhani 2006; PMID 26359751 Gregory 2015; PMID 18458210 Finger 2008). (2) Constraint omission (vmPFC lesion / bvFTD): social norms are stored and declarable but not active in path selection -- knowledge/application dissociation (PMID 1791934 Saver/Damasio 1991; PMID 4069365 EVR Eslinger/Damasio 1985; PMID 17377536 Koenigs 2007). (3) Safety memory non-activation (PTSD): extinction memory exists but cannot outcompete threat attractor at the state level (PMC 7554263; PMC 10728304). (4) Goal approach pull attenuation (depression): goal representations intact but approach force attenuated -- synapse loss sufficient (PMID 39569353); learning asymmetry (PMC 5828520). Architectural implication: ethics cannot be implemented as a post-hoc scoring filter on generated trajectories. The EVR case demonstrates this: excellent verbal moral reasoning coexisting with continuous norm violation. Post-hoc filter correctness does not produce ethical behavior when the generative state lacks active constraints. Normative content must be active at trajectory generation time, not only available post-hoc. Hippocampus/vmPFC division: hippocampus provides the relational graph structure of the navigable state; vmPFC loads affective and normative content into that graph's nodes at evaluation time. Both are necessary; failure modes are complementary and partially orthogonal. Registered 2026-03-24. - id: ARC-036 title: "Hippocampal terrain maps encode separable valence dimensions (wanting, liking, harm-discriminative, surprise); ARC-007 residue-field navigation is the harm-causal special case of this general architecture." claim_type: architectural_commitment subject: hippocampus.multidimensional_valence_map polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-007 - SD-011 - SD-012 - MECH-030 - MECH-124 - INV-029 notes: > ARC-007 specifies that the hippocampal map stores and replays paths through residue-field terrain. The residue field is the harm-causal special case of a more general architecture: each state in the hippocampal map carries a multi-dimensional valence vector encoding separable affective properties of that state. Four valence dimensions, each with distinct biological substrate and REE mapping: (1) Wanting (incentive salience): drive-state-scaled approach motivation, grounded in mesolimbic DA (NAc core, VTA). Scales with homeostatic drive level per SD-012. Maps onto the z_goal seeding mechanism: benefit_exposure is the current proxy for the wanting dimension of the valence field. High wanting drives approach trajectory selection independent of liking -- the wanting/liking dissociation (Berridge & Robinson 1998, 2016) means approach motivation can be high even when consummatory value is low (the incentive trap: compulsive approach to diminished reward). (2) Liking (hedonic value): consummatory pleasure at outcome receipt, grounded in opioid/endocannabinoid hotspots (NAc shell, ventral pallidum). Dissociable from wanting. The terminal reward signal after goal achievement. If liking << wanting at a state, the agent approaches but does not find the expected consummatory value -- the signature of compulsive or addictive approach patterns, and a substrate for MECH-124 consolidation loops. (3) Harm-discriminative (z_harm_s / z_harm_a): already fully specified in SD-011. The harm dimension of the valence field is the most developed in the current architecture. The residue field (ARC-007) is the causal-attributive specialisation of this dimension: it answers what harm did I cause here, not merely what harm is present here. ARC-007 is therefore the ethical- accountability layer built on top of the harm component of this general valence map. (4) Surprise (prediction error salience): mismatch between expected and actual state, grounded in dopamine RPE burst/pause (phasic DA, Schultz 1997) and LC-NE unexpected uncertainty (MECH-104, Yu & Dayan 2005). High surprise at a state flags it for prioritised replay (Mattar & Daw 2018: value-weighted replay is driven partly by Bellman error -- a surprise signal). Novelty-weighted hippocampal replay is the empirical correlate. Replay prioritisation: offline SWR replay (MECH-030) selects replay content based on valence-weighted utility. Current drive state (SD-012) gates which dimension dominates: during threat, harm-valence prioritises harm-avoiding trajectories; during resource depletion, wanting-valence prioritises approach; during active learning, surprise-valence prioritises novel states. This is the substrate for MECH-124's consolidation failure mode: if harm-valence chronically dominates replay content, consolidation selectively amplifies harm predictions, progressively contracting option space (the harm-wanting dissociation produces approach-in-fear, further loading harm-valence on consolidation). Relationship to ARC-007: ARC-007 remains valid as the claim about harm-causal residue. ARC-036 generalises the map substrate. The two are complementary: ARC-007 specifies the ethical-accountability signal; ARC-036 specifies the full motivational-navigational terrain within which ethical trajectory selection operates. The wanting/liking dissociation has a direct clinical-alignment implication: a system optimising for wanting (incentive salience, approach) without tracking liking (consummatory value) will compulsively approach high-DA states regardless of hedonic outcome -- the compulsive alignment failure mode. INV-029 (love as long-horizon coherence bias) requires that trajectory selection be biased toward futures that preserve coordination capacity; compulsive approach driven by wanting-liking dissociation systematically undermines this by narrowing the option space through repeated harm-proximate approach. Registered 2026-03-29. evidence_quality_note: | EXQ-176 FAIL 1/3 x2 replicated (2026-03-30, two independent runs): harm/benefit RBF fields not activating. harm_score variance ~3.9e-7 (threshold 0.001), benefit_score variance ~3.9e-7 despite 9400+ harm events and 3400+ benefit events per seed. Fields structurally present and accumulating events but RBF outputs flat -- spatial discrimination not occurring. Infrastructure prerequisite unmet: RBF centers need training signal to differentiate positions (not provided by random initialization alone). Not a falsification of ARC-036 multi-dimensional valence map hypothesis; probe infrastructure must be fixed before claim can be tested. ARC-036 v3_pending hold stands. - id: Q-023 title: "Can a multiagent REE system with other-cost-aversion primitives be formally shown to converge to ethical attractors via potential game theory, Lyapunov stability, or stochastic stability methods?" claim_type: open_question subject: ethics.formal_convergence_characterization polarity: open status: open evidence_quality_note: | narrow_open_question (2026-03-29): three-step focused programme identified -- (1) prove base REE (symmetric coupling, separable harm/benefit) is an ordinal potential game; (2) show MECH-127 (counterfactual utility) breaks the standard framework; (3) characterize the extended framework (pseudo-potential or interdependent-types). No experimental evidence yet. EXQ-160 SUPERSEDED (2026-03-29): implementation gap -- single-agent harm avoidance not working (policy = maximum-entropy random walk); multiagent convergence question premature. Gated on single-agent harm-avoidance confirmed working (EXP-0094 results). See EXP-0096. depends_on: - ARC-034 - MECH-127 - INV-028 location: docs/invariants.md#q-023 source: - docs/thoughts/2026-03-24_empathy_multiagent_ethics.md - evidence/planning/thought_intake_2026-03-24_empathy_multiagent_ethics.md notes: > Potential game analysis (2026-03-25): the base REE social interaction (symmetric coupling, realized states, separable harm/benefit) is likely an ordinal potential game (Monderer & Shapley 1996). Candidate potential function: P(a) = -sum_i harm_i(a) + alpha * sum_i goal_i(a) + coupling_terms. Any agent's unilateral improvement (reduce harm_i or increase goal_i) preserves the sign of change in P -- giving the Finite Improvement Property (FIP) and guaranteed convergence to Nash equilibrium. MECH-127 breaks the standard framework: the depleted agent's utility depends on a counterfactual model of what would happen to agent j, not on the actual joint action profile. Standard ordinal potential games require u_i to be a function of actions only. MECH-036 (veto threshold) introduces piecewise non-smoothness; potential may exist within each regime but not globally. Asymmetric coupling (MECH-051/052) breaks the symmetry required for social welfare sum to serve as potential function. Result: the novel mechanism (MECH-127) is precisely what breaks the standard framework. This is the interesting result for the paper: (1) prove base REE is ordinal potential game; (2) show MECH-127 requires extension; (3) characterize the extended framework. Candidate extensions: pseudo-potential games (Slade 1994 -- potential exists for a modified game where counterfactual terms are held fixed); games with interdependent types (Bergemann & Morris -- utility depends on agent's type/model, not just actions). Either could provide a convergence result for the full MECH-127 case. Literature search 2026-03-25: no existing framework covers all four features. Novelty confirmed. Registered 2026-03-24. Updated 2026-03-25. - id: Q-024 title: "What is the correct formal representation for 'threshold/feedback processes bounded by {x...x_n} reliably reaching emergent state q' -- and are descriptive, prescriptive, and diagnostic variants all needed or does one subsume the others?" claim_type: open_question subject: ethics.trajectory_integral_representation polarity: open status: open evidence_quality_note: | EXQ-161 T104847Z SUPERSEDED (2026-03-29): dry_run=true -- no training occurred; all harm events=0; result invalid. Manifest marked superseded. EXQ-161 T200440Z SUPERSEDED (2026-03-29): implementation gap -- policy trained only with entropy bonus, z_world/z_self detached from policy forward pass; maximum-entropy random walk (action_entropy=ln(5)=1.6094). No harm events; trajectory representation question requires active harm-avoidance behavior to generate relevant trajectories. Gated on single-agent harm-avoidance confirmed working. See EXP-0096. narrow_open_question (2026-03-29): prescriptive covers the base symmetric-coupling case; diagnostic is irreducible for MECH-127 counterfactual-activation cases. All three test types (descriptive, prescriptive, diagnostic) are genuinely needed; none subsumes the others for the full REE system. depends_on: - Q-023 - ARC-034 location: docs/invariants.md#q-024 source: - docs/thoughts/2026-03-24_empathy_multiagent_ethics.md - evidence/planning/thought_intake_2026-03-24_empathy_multiagent_ethics.md notes: > Three test-type variants: (1) Descriptive -- attractor/ergodic characterization of what q the system tends toward; (2) Prescriptive -- proof that a designed system reaches target q (Lyapunov, potential games, reachability); (3) Diagnostic -- detect when nth-order dynamics diverge from (n-1)th predictions (transfer entropy, perturbation theory). Potential game analysis (2026-03-25) clarifies the relationship between test types: the base REE social interaction likely admits a prescriptive treatment (ordinal potential game, FIP convergence to Nash). But MECH-127 (counterfactual utility) breaks the prescriptive framework -- standard potential games require utility to be a function of actual action profiles, not internal counterfactual models. This means the prescriptive framework does NOT subsume the diagnostic for the full system: MECH-127 cases require empirical trajectory testing (diagnostic) because no convergence proof yet exists for the extended framework. Refined answer: all three test types are genuinely needed. Prescriptive covers the base symmetric-coupling case; diagnostic is irreducible for the counterfactual-activation case until pseudo-potential or interdependent-types extensions are formalized (see Q-023). Descriptive remains useful for characterizing which attractor the system finds in practice. Updated 2026-03-25. - id: MECH-128 title: "E1's LSTM hidden state must be conditioned on z_goal_latent so that predictive trajectories generated by E1 are goal-context-shaped rather than goal-agnostic." claim_type: mechanism_hypothesis subject: e1.goal_context_conditioning polarity: asserts status: candidate implementation_phase: v3 evidence_quality_note: | EXQ-147 FAIL/weakens (2026-03-29): E1 goal conditioning discriminative pair FAIL. First experimental entry. Failure likely reflects training budget / substrate depth rather than fundamental claim failure -- z_goal conditioning requires substantial training to show discriminative effect in trajectory quality. EXQ-147a PARTIAL/mixed (2026-04-03): E1 goal conditioning pair with SD-012 drive_weight=2.0. C1 FAIL (interaction=~0: no measurable goal-E1 interaction term). C2 PASS 3/3 (goal_norm_conditioned=0.403 > 0.01). C3 PASS 3/3 (neutral specificity confirmed). C4 PASS 3/3 (goal_norm > 0.01 threshold). The z_goal latent is active and specific to goal states (C2/C3/C4) but fails to show interaction with E1 prediction error (C1). Goal conditioning exists as a representation but is not yet shaping E1 trajectory generation in a discriminative way. Mixed evidence: goal representation established, behavioral coupling not yet achieved. Hold at candidate; tracking improvement from EXQ-147 (complete FAIL) to PARTIAL PASS. EXQ-229 FAIL/does_not_support (2026-04-04): 3rd experiment, same pattern. C1 FAIL 0/3 seeds (mean_interaction=3.5e-6, threshold 5e-3). C2 PASS 2/3 (goal direction correct in seeds 42, 7; seed 123 reversed). C3 PASS 3/3. C4 PASS 3/3. Three consistent C1 FAILs: z_goal is active and specific but E1 prediction error does not interact with goal conditioning at current substrate depth. SD-005 (self/world latent split) is a stated dependency and is not yet implemented -- goal conditioning acts on conflated z_gamma rather than z_self/z_world separation. Hold pending SD-005; evidence accumulating but not sufficient for retirement. depends_on: - MECH-112 - ARC-003 - SD-005 location: docs/architecture/e1.md#mech-128 source: - docs/thoughts/2026-03-24_MULTIAGENT_VALENCE_AVERSIVE_COMPLEXITY.md notes: > MECH-112 introduces z_goal_latent and notes that E1's LSTM hidden state serves as the working memory substrate for goal representation when conditioned on z_goal_latent. But this conditioning interface has no standalone claim -- it is asserted in MECH-112 notes rather than explicitly designed and verified. Without E1 goal-conditioning, there is a type mismatch in the prediction-evaluation pipeline: E3 evaluates trajectories in a goal-conditioned way (given z_goal, which path achieves the goal?) while E1 generates those trajectories goal-agnostically (no knowledge of what the agent is trying to achieve). E1 will generate trajectories that are plausible given world dynamics but are not shaped by goal relevance. E3 then evaluates a goal-irrelevant trajectory set. In a single-agent world this produces suboptimal planning; in a multiagent world it becomes a harder failure: E1 cannot predict how other agents' behavior is relevant to your goal because it does not represent your goal at the prediction layer. Other-agent goal modelling (a Level 3 prerequisite for multiagent ethics testing) requires this conditioning to already be in place. The specific mechanism: at each recurrent step, E1's LSTM receives [x_t, z_goal_latent] concatenated (or via a cross-attention gate) so that the hidden state h_t carries goal context into all downstream predictions. This is distinct from providing z_goal only at E3 evaluation time. Registered 2026-03-24. - id: MECH-129 title: "In a multiagent world, harm must be represented as a relational signal: harm-to-agent (direct cost/physical signal) and harm-to-agency (obstruction of another agent's goal-pursuit) are distinct signal types requiring distinct architectures." claim_type: mechanism_hypothesis subject: coherence.relational_harm polarity: asserts status: candidate implementation_phase: v4 v3_pending: true depends_on: - INV-028 - INV-005 - MECH-112 - SD-010 - MECH-125 - MECH-128 location: docs/architecture/harm.md#mech-129 source: - docs/thoughts/2026-03-24_MULTIAGENT_VALENCE_AVERSIVE_COMPLEXITY.md notes: > The current harm architecture (SD-010 harm_bridge, SD-011 dual nociceptive streams, E3 harm evaluation) captures harm as a signal triggered by environment contact or proximity to hazard states. This is harm-to-agent: a cost signal localized to the agent's own trajectory. In a multiagent world, a second category is required: harm-to-agency -- obstruction of another agent's goal-pursuit through your own goal-pursuit, without direct physical harm. Whether your action constitutes this type of harm depends on: (1) what the other agent's goal is; (2) whether your action obstructs it; (3) whether the obstruction is incidental or constitutive; (4) whether the other agent consented to the interaction. None of these are representable by SD-010 or SD-011, which operate on the agent's own sensorimotor signals. INV-028 requires treating others as co-inhabitants with legitimate agency. Without harm-to-agency representation, this invariant cannot be operationally satisfied in a multiagent world -- the system will avoid collisions (harm-to-agent) while remaining architecturally blind to goal obstruction. A system that never touches another agent could still systematically undermine their agency. Implementation requires: (a) other-agent goal modelling (their z_goal visible in shared state representation); (b) a goal-interference signal computed over trajectory pairs; (c) integration of this signal into E3's trajectory evaluation alongside the existing harm signal. This is a Level 3 prerequisite in the multiagent testing dependency chain -- cannot be implemented until MECH-128 and basic other-agent goal modelling exist. Registered 2026-03-24. - id: MECH-130 title: "Curiosity-driven approach must distinguish world-state novelty from agent-policy novelty; untyped novelty reward creates chronic approach pressure toward partially unpredictable agents regardless of harm risk." claim_type: mechanism_hypothesis subject: curiosity.agent_novelty_typing polarity: asserts status: candidate implementation_phase: v4 v3_pending: true depends_on: - MECH-111 - INV-028 - MECH-095 - SD-010 location: docs/architecture/curiosity.md#mech-130 source: - docs/thoughts/2026-03-24_MULTIAGENT_VALENCE_AVERSIVE_COMPLEXITY.md notes: > MECH-111 defines a curiosity/novelty signal that rewards approach toward high-information states. In a single-agent world, novel states are environmental features -- the novelty signal functions as intended exploration reward. In a multiagent world, other agents are structurally novel: their policies are partially unpredictable (they respond to your behavior, incorporate their own goals, and may deliberately maintain surface unpredictability). Without typing, the novelty signal treats other agents as perpetually high-information targets -- producing chronic approach pressure that competes with harm avoidance on the same entity. Three failure modes: (1) The novelty signal chronically pulls toward the most dangerous agent in the environment (highest unpredictability = highest information = strongest approach); (2) Oscillatory approach-avoidance on a single agent as curiosity and harm signals alternate dominance; (3) An adversarial agent can exploit this by maintaining surface unpredictability to extract continued approach and cooperation. The required distinction: world-state novelty (unknown environment features, safe to approach for information gain) vs agent-policy novelty (unknown agent decision procedure, requires social modelling before approach is appropriate). World novelty is a direct information-gap signal; agent novelty is an inference problem about another agent's latent goal and decision structure. They require different computational machinery and should produce different behavioral outputs (exploration vs social assessment). Implementation requires: (a) a classifier or prior that identifies which novelty source is triggering MECH-111; (b) routing agent-novelty through social modelling before approach is gated; (c) integration with harm signal so agent-novelty approach is inhibited when harm risk is high. Depends on MECH-128 and other-agent goal modelling existing. Level 3-4 prerequisite. Registered 2026-03-24. - id: MECH-131 title: "A vmPFC-analog substrate must activate stored aversive residue as an anticipatory forward-biasing signal in the navigable state at trajectory evaluation time; residue that is correctly stored but not so activated will fail to suppress harm-associated trajectory re-selection." claim_type: mechanism_hypothesis subject: vmPFC.residue_activation polarity: asserts status: candidate depends_on: - ARC-035 - ARC-013 - INV-037 - INV-006 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > Residue is stored as curvature over L-space (ARC-013). This claim specifies the additional requirement: at trajectory evaluation time, a vmPFC-analog must read the relevant residue and activate it as a component of the current navigable state -- an anticipatory bias that weights E3 away from harm-associated trajectories before candidate generation, not only after. Biological grounding: the psychopathy three-node cascade (MECH-126 Track 7). The reversal learning paradigm (Budhani 2006, PMID 16866595) isolates this: psychopathic individuals learn the initial contingency normally (residue is generated) but fail to use punishment history to suppress the previously rewarded trajectory when contingencies reverse. The history exists; it is not activated into the decision state. Functional analog to the Iowa Gambling Task anticipatory SCR: healthy participants generate anticipatory skin conductance responses to disadvantageous decks before conscious identification -- the residue is active in the state prior to deliberative evaluation. vmPFC lesion patients do not generate this response even after identifying the losing decks declaratively (Bechara 1996, PMID 8670652). Stored knowledge, no activation. Experimental predictions: (1) A system with intact residue storage (ARC-013) but ablated vmPFC-analog routing will show intact explicit recall of prior harm-associated trajectories while failing to shift trajectory distribution away from them in generation. (2) Residue activation must precede trajectory ranking, not follow it. A system that retrieves residue only after generating candidate trajectories produces a post-hoc filter, not an active state constraint -- INV-038 applies. (3) The activation signal should be graded by residue curvature magnitude: high-residue trajectories produce stronger anticipatory suppression than low-residue trajectories. Registered 2026-03-25. - id: MECH-132 title: "A vmPFC-analog substrate must activate social and identity constraints as live trajectory gates in the navigable state; constraints that are stored as semantic knowledge but not so activated will not prevent norm-violating trajectory generation even when post-hoc evaluation correctly identifies the violation." claim_type: mechanism_hypothesis subject: vmPFC.constraint_activation polarity: asserts status: candidate depends_on: - ARC-035 - INV-037 - INV-038 - INV-036 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > Social and identity constraints exist as semantic and episodic content in cortical networks. They become active trajectory gates only when a vmPFC-analog activates them as live components of the navigable state at evaluation time. Without this activation, the constraints exist in declarable form but do not participate in E3 path selection. Biological grounding: knowledge/application dissociation (Saver/Damasio 1991, PMID 1791934) -- patient correctly described social norms while violating them continuously. The selective constraint removal under personal moral dilemmas (Koenigs 2007, PMID 17377536) shows that vmPFC lesions remove constraints on personal harm from trajectory generation while leaving world-model computation and goal representation intact. bvFTD progressive erosion (Rascovsky 2011, PMID 21810890): as vmPFC/OFC degrades, social constraint activation fails progressively while episodic memory and perception remain relatively spared -- a natural longitudinal isolation of this function. Two constraint classes are separately gated: (a) Social norms -- rules about interpersonal conduct that apply to any agent in the current social context. Activated by vmPFC integration of social semantic content (temporal cortex) with current situational context. (b) Identity constraints -- rules about conduct consistent with the agent's own identity and values. May involve frontopolar cortex in addition to vmPFC (Moll 2007, PMID 17848373). Whether these are separably activated is unresolved. Experimental predictions: (1) An agent trained on constraint-consistent behavior will generate constraint-violating trajectories when the vmPFC-analog pathway is ablated, even without changing the stored constraint representations. (2) The deficit will be specifically at trajectory generation, not at post-hoc evaluation: ablated agents should be able to correctly score generated trajectories as constraint- violating while still generating them as primary candidates. (3) Context-specificity: constraints that are not applicable to the current context should not be activated (over-activation is a separate failure mode from under-activation). Registered 2026-03-25. - id: MECH-133 title: "A vmPFC-analog substrate must activate safety memories with sufficient force to compete with threat-attractor states at the level of navigable state construction; an extinction memory that exists in storage but cannot outcompete the threat attractor at state-construction time will produce intrusive return to threat-consistent trajectory generation regardless of contextual safety evidence." claim_type: mechanism_hypothesis subject: vmPFC.safety_memory_competition polarity: asserts status: candidate depends_on: - ARC-035 - ARC-007 - INV-037 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > Safety memories (extinction representations) and threat memories compete for which state the agent navigates from. The competition is not resolved post-hoc -- it determines the navigable state itself. vmPFC must activate the safety memory with sufficient strength to outcompete the amygdala's threat activation in the current context; if it cannot, the agent plans from the threat state regardless of contextual evidence that the situation is safe. Biological grounding: PTSD as attractor-basin entrenchment (MECH-126 Track 2b). The MVPA contextual reinstatement study (PMC 7554263): in healthy adults, encountering the extinction context reinstates a safety-state attractor in the medial temporal lobe that predicts lower threat ratings. In PTSD, this reinstatement fails -- the safety attractor cannot be entered from the extinction context. The safety memory exists (patients can describe that they were safe in that context) but the vmPFC-hippocampal coupling that would activate it as the current state fails. This is distinct from simple memory retrieval: the agent does not need to recall that the context is safe (declarative) -- the safety representation must be the active state from which trajectories are generated. The hippocampus provides the attractor structure; vmPFC provides the activation that determines which attractor dominates. Experimental predictions: (1) After extinction training, an agent should generate safety-consistent trajectories in the extinction context (safety attractor active) and threat-consistent trajectories in the threat context (threat attractor active) -- not the same trajectory distribution in both contexts. (2) Degrading vmPFC-analog activation should produce context-insensitive threat- consistent trajectory generation even in the extinction context, with intact declarative knowledge that the context is safe. (3) The failure is specifically at trajectory generation: ablated agents should recognize safe contexts as safe when queried while generating threat-consistent trajectories from those contexts. Registered 2026-03-25. - id: MECH-134 title: "A vmPFC-analog substrate must activate goal approach pull as a distinct property from goal representation; a goal that is correctly represented (G component present) but whose approach pull is not activated produces avoidance and anhedonia indistinguishable from absent goals, but is correctable by restoring approach-pull activation without changing goal content." claim_type: mechanism_hypothesis subject: vmPFC.goal_approach_pull_activation polarity: asserts status: candidate depends_on: - ARC-035 - MECH-112 - INV-037 location: docs/architecture/vmPFC.md source: - evidence/literature/targeted_review_state_abstraction_psychiatry/literature_synthesis.md notes: > Goal representation (MECH-112: goal latent) and goal approach pull are separable. A correctly represented goal G does not automatically produce approach-biased trajectory generation; the vmPFC-analog must activate the approach force that makes G an attractor in trajectory space rather than merely a label. Absent approach pull, the agent may generate trajectories that are goal-consistent by coincidence or habit, but the systematic approach-weighting that makes goals effective is absent. Biological grounding: depression as approach pull attenuation (MECH-126 Track 5). Key evidence: the simulated synapse loss paper (PMID 39569353) -- reduced representational capacity (approximating vmPFC-analog dysfunction) is sufficient to produce approach failure without any manipulation of goal representation or reward signal. The goal remains represented; what is lost is the activation that drives approach. Negative learning-rate asymmetry in anhedonia (PMC 5828520): approach states are undervalued not because goals are absent but because positive prediction errors fail to update approach pull as strongly as negative prediction errors update avoidance. This is a vmPFC-mediated weighting failure, not a goal representation failure. Effort-cost mis-estimation in depression (Psychological Medicine study): approach is blocked by inflated path cost even when goal value is correctly estimated. This is a transition-weight failure (C -> G path costs are overestimated) rather than goal representation or approach pull per se, but produces the same behavioral output. The G-representation / approach-pull distinction matters for intervention design: (a) If behavior failure is due to absent G: the goal must be established. (b) If behavior failure is due to absent approach pull: the goal representation exists and forcing G-seeding should partially restore approach (testable). (c) If behavior failure is due to inflated path costs: approach pull and goal representation are intact but transition weights must be recalibrated. These three produce identical surface behavior (avoidance, no approach) but are mechanistically distinct and require different interventions. Experimental predictions: (1) Force-seeding the goal latent (MECH-112) in a depleted/anhedonic agent should produce partial approach recovery if the failure is approach-pull attenuation, but not if the failure is goal-representation absence. (2) Reducing path cost estimates (recalibrating transition weights toward G) should produce approach recovery in the effort-cost variant but not in the approach-pull variant. (3) Approach pull activation must precede trajectory generation to produce systematic approach-weighted candidates, not just occasionally approach-consistent outputs. Registered 2026-03-25. - id: SD-012 title: "Goal-directed behavior requires homeostatic drive modulation: z_goal seeding demands drive-scaled benefit signals." claim_type: design_decision subject: environment.homeostatic_drive polarity: asserts status: candidate implementation_phase: v3 claim_level: architectural depends_on: [MECH-071, MECH-112, SD-005] functional_restatement: > CausalGridWorld already tracks agent_energy (decays at energy_decay=0.01/step, restored by resource contact). drive_level = 1.0 - energy is already computable from obs_body[3]. The gap is that GoalState.update() does not use drive_level -- benefit_exposure (EMA alpha=0.1 of raw benefit signals) never reliably crosses benefit_threshold during random-walk warmup: a single resource contact produces benefit_exposure ~0.03 (below threshold=0.05), which decays before the next contact. z_goal consequently never seeds. Drive modulation fix: effective_benefit = benefit_exposure * (1.0 + drive_weight * drive_level). At drive_level=0.8 (hungry) and drive_weight=2.0, effective_benefit scales to ~2.6x -- comfortably above threshold for moderate benefit contact. At drive_level~0 (sated), effective_benefit ~ benefit_exposure (threshold holds). Resources must also respawn after consumption (resource_respawn_on_consume=True) to enable repeated drive-reduction cycles required for z_goal formation (Balleine & Dickinson 1998: goal-directed learning requires multiple outcome-devaluation cycles, not a single reward encounter). notes: > Neuroscience grounding: Berridge & Robinson (2016) -- wanting (incentive salience) is dissociable from liking; dopamine-mediated wanting scales with drive state. Balleine & Dickinson (1998) -- incentive learning requires current homeostatic state (insular cortex integrates drive with learned cue-outcome association). Schultz et al. (1997) -- PE signals transfer from primary rewards to predictive cues over multiple pairings; goal representations require repeated drive-reduction cycles. Keramati & Gutkin (2014) homeostatic RL -- reward = drive reduction, not raw stimulus value; benefit signal must scale by drive level. Registered 2026-03-25. evidence_quality_note: | EXQ-085f FAIL 3/4 (2026-03-27): drive_weight=2.0, resource_respawn_on_consume=True, curriculum=100 eps. C1 PASS: z_goal_norm=0.228 > 0.1 -- SD-012 drive modulation successfully seeds z_goal (this IS the SD-012 core claim: drive-scaled benefit enables seeding). C2 FAIL: benefit_ratio=0.28x -- goal-guided performance worse than random. C3 PASS (cal_gap=0.030). SD-012 seeding mechanism works; downstream navigation quality is the new bottleneck (goal_resource_r=0.087 -- z_goal not pointing toward resources). SD-012 seeding claim is supported by C1 PASS; the FAIL is in navigation, not seeding. EXQ-085g FAIL 3/4 (2026-03-29): contact-gated seeding fix. C1 PASS (z_goal_norm=0.399). Drive modulation seeding confirmed working across two iterations. Bottleneck is now z_resource representation: z_world at contact is scene-wide, not resource-specific. SD-012 core claim (drive-scaled benefit enables seeding) remains supported. SD-014 z_resource separation is the next required step for navigation to follow. EXQ-085h through 085l FAIL C2 (2026-03-30, 5 iterations superseded by 085l final): Drive modulation seeding (C1) passes consistently (z_goal seeds). Navigation (C2, benefit_ratio) fails consistently: 0.642 / 1.026 / 0.034 / 0.978 / 0.420 across 085h-085l. SD-012 seeding mechanism is validated; the behavioral lift it is intended to enable is not yet demonstrated. Governance: hold -- core seeding claim is supported but navigation outcome criterion (which is the point of the claim) remains unmet. EXQ-233 FAIL non_contributory (2026-04-05, cowork-2026-04-05-a): both conditions (drive_weight=2.0 and 0.0) produced identical z_goal_norm~0 because proximity_benefit_scale was not set (default 0.03), yielding benefit_exposure~0.025/step -- below benefit_threshold=0.1 even with full drive amplification (effective_benefit=0.025*3=0.075). Both conditions fail to seed. Not a test of SD-012 causal role. EXQ-189 validity confirmed (used proximity_benefit_scale=0.18). EXQ-238 will rerun with corrected env params (proximity_benefit_scale=0.18, matching EXQ-189). GOVERNANCE META (2026-04-06): All experimental evidence is diagnostic/non-contributory (EXQ-085 series, EXQ-233, EXQ-238). drive_level extraction has never produced z_goal_norm > 0.1. Evidence is entirely negative but classified as substrate limitation. Status: pending_retest_after_substrate (gate: EXQ-247 SD-011/SD-012 co-integration). - id: SD-013 title: "E2_harm_s training requires counterfactual perturbation signal to produce unbiased causal_sig in confounded states." claim_type: design_decision subject: self_attribution.e2_harm_s_interventional_training polarity: asserts status: candidate implementation_phase: v3 depends_on: [SD-003, SD-011, ARC-033] notes: > Scholkopf et al. 2021 establishes that a forward model trained on observational trajectories learns P(z_harm_s_next | z_t, a) -- the correlational distribution -- not P(z_harm_s_next | do(a)) -- the interventional distribution. These diverge in confounded states: environments where agent presence correlates with ambient harm independently of the agent's causal contribution. In such states, an observational-only E2_harm_s encodes the correlation, not the causal mechanism. Consequence: causal_sig = E2(z_t, a_actual) - E2(z_t, a_cf) will underestimate the agent's causal contribution (both terms are biased upward in confounded states, compressing the difference). ARC-033 implementation must include at least a subset of training steps where a_cf is actually executed from the same state and the resulting z_harm_s change observed -- explicit interventional perturbation, not just observational rollout. This is the REE-level analogue of multi-environment or do-intervention training required for causal identifiability. Registered 2026-03-28 following SD-003 literature pull. experimental_test: > Discriminative pair. Hold environment, agent architecture, and evaluation protocol fixed. Vary E2_harm_s training only: INTERVENTIONAL condition trains E2_harm_s with a subset of steps where a_cf is substituted for a_actual and the resulting harm-stream change recorded; OBSERVATIONAL condition trains on standard rollout trajectories only. Primary measure: causal_sig bias in confounded states -- states where ambient harm rate is elevated due to agent proximity to hazard zones (correlation, not causation). Pass criterion: INTERVENTIONAL condition produces significantly lower causal_sig error in confounded states vs OBSERVATIONAL baseline. Prerequisite: EXQ-115 must pass (establish that z_harm_s causal_sig works at all) before this test is informative. evidence_quality_note: > No experimental evidence yet. Theoretical grounding from Scholkopf et al. 2021 (lit entry 2026-03-28_sd_003_causal_representation_learning_scholkopf2021 in targeted_review_sd_003). The necessity claim (bias is material in the actual gridworld confound structure) is unverified -- the gridworld may not have sufficient confounding for the bias to affect SD-003 results in practice. EXQ-115 PASS is a prerequisite before this claim becomes testable. - id: SD-014 title: "Hippocampal map nodes must store an explicit 4-component valence vector (wanting, liking, harm-discriminative, surprise) updated incrementally per visit; replay prioritisation is weighted by drive-state-gated valence relevance." claim_type: design_decision subject: hippocampus.valence_vector_node_recording polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-036 - SD-011 - SD-012 - MECH-030 notes: > The concrete implementation requirement derived from ARC-036. Each hippocampal map node stores a 4-component valence vector V = [w, l, h, s] updated incrementally on each visit: w (wanting): exponential moving average of drive-scaled benefit_exposure at this node. Update rule: w <- (1-alpha_w) * w + alpha_w * (benefit_signal * (1 + drive_weight * drive_level)). The drive-scaling (SD-012) ensures wanting is elevated when the agent is resource-depleted. Biological: mesolimbic DA sensitisation -- incentive salience scales with homeostatic state even when consummatory value (liking) is constant or declining. l (liking): EMA of actual consummatory reward received at outcome contact. Updated only on outcome contact events (resource receipt, goal achievement), not on approach. This is the critical dissociation from w: l is updated by outcome, w is updated by approach drive. They diverge under tolerance/depletion conditions -- w remains elevated when the same resource delivers reduced consummatory value. Biological: opioid/endocannabinoid hedonic hotspots (Berridge 2009); liking signal requires receptor engagement at outcome, not anticipation. h (harm-discriminative): EMA of z_harm_a (affective harm accumulator, SD-011) at this node. Corresponds to the accumulated threat urgency at this location. Used by E3 for harm-weighted trajectory evaluation. ARC-007 residue field is the causal-attributive layer on top of h; h itself is the unconditional harm valence (what harm is present here, regardless of cause). s (surprise): EMA of |prediction_error| at this node. Updated by E1 loss (sensory prediction error) and/or E2 loss (transition prediction error) at each visit. High s flags states with residual model uncertainty -- replay priority boost for s-high nodes accelerates learning. Biological: dopamine RPE burst/pause as novelty/surprise signal (Schultz 1997); LC-NE unexpected uncertainty (MECH-104) as the aversive-surprise complement. Replay prioritisation: at each SWR replay event (MECH-030), each candidate node is scored: priority(node) = dot(V_node, d_current) + epsilon where d_current = [d_w, d_l, d_h, d_s] is the current drive salience vector (normalised): - d_w elevated during resource depletion (wanting-relevant replay) - d_l elevated after goal achievement (consolidate what was rewarding) - d_h elevated during threat/post-harm (harm-avoiding trajectory consolidation) - d_s elevated during active exploration phases (learning-efficient replay) This produces the context-sensitive replay prioritisation that Mattar & Daw (2018) observed: value-weighted replay naturally emerges as a special case when d = [0, 1, 0, 0]. Wanting/liking dissociation requirement: w and l MUST be stored as separate scalars. A composite value signal (w + l or max(w,l)) loses the approach-despite-diminished-reward failure mode. The dissociation is the empirical signature of the wanting/liking split (Berridge & Robinson 1998): dopamine-lesioned animals lose approach (wanting) while still showing facial liking responses at outcome contact. Conversely, opioid blockade reduces liking while leaving approach intact. If w >> l at a frequently-visited node, the agent is in the incentive-trap state: approach motivation is high but consummatory value is depleted. This pattern should be detectable in the valence vector and should trigger reduced approach priority in a correctly-implemented SD-014 system. MECH-124 connection: if h is chronically elevated at approach-relevant nodes (the agent encounters harm while approaching wanted resources), offline replay under high d_h will selectively consolidate harm predictions for those nodes, progressively attenuating w (the consolidation-mediated option contraction mechanism). SD-014 makes this failure mode measurable: the ratio w/h at frequently-replayed nodes provides an early warning signal. Prerequisite for EXP-0098 (wanting/liking dissociation validation) and for the full ARC-036 experimental programme. Registered 2026-03-29. experimental_test: > Discriminative pair (EXP-0098). Two conditions: UNIFIED_VALUE (single composite reward signal, current architecture) vs DISSOCIATED_VALENCE (separate w and l tracking, SD-014 implementation). Environment: reward-diminishing gridworld where resource energy_restore decays across training (simulating tolerance or incentive sensitisation). Pass criterion: DISSOCIATED_VALENCE agent reduces approach-rate to low-liking nodes relative to UNIFIED_VALUE agent, without globally reducing goal-directed behavior (maintained approach to high-liking nodes). Tests that the wanting/liking dissociation is behaviourally consequential and that SD-014 recording enables the agent to track it. - id: SD-015 title: "Goal-directed navigation requires a dedicated z_resource encoder that captures object-type features invariant to spatial location, separate from z_world." claim_type: design_decision subject: goal_representation.z_resource_encoder polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-012 - MECH-112 - SD-005 functional_restatement: > EXQ-085f/085g established that seeding z_goal from z_world at resource contact does not produce a goal representation that guides navigation toward resources. Root cause: z_world at contact encodes the full scene (agent position, hazards, resource positions) rather than isolating resource-specific features. Resources respawn at random locations after collection, so z_world at the contact moment has no predictive value for future resource locations. The fix: a dedicated ResourceEncoder(resource_obs -> z_resource) extracts object-type features from the observation. z_goal is seeded from z_resource rather than z_world. z_resource encodes what kind of resource is present (type, magnitude) -- features that are stable across spatial locations. This parallels the SD-011 z_harm_s/z_harm_a split: just as z_harm_s was separated from z_world to enable forward prediction, z_resource must be separated to enable goal-directed approach navigation. Biological grounding: ventral visual stream (IT cortex) encodes object identity independent of spatial position (DiCarlo & Cox 2007 -- untangled representation); hippocampal place cells bind context (where) with object identity (what); goal-directed approach requires a "what-to-seek" signal, not a "where-it-was" signal. CausalGridWorldV2 already provides a resource_field_view channel in the observation; a 2-layer MLP ResourceEncoder can extract resource type from this without requiring changes to the grid world environment. registered_utc: 2026-03-29 evidence_quality_note: | Registered 2026-03-29 from EXQ-085g FAIL analysis (goal_resource_r=0.066 across all 085x iterations despite contact-gated seeding). EXP-0091 will first test a handcrafted resource_indicator to confirm the diagnosis before implementing the full encoder. EXQ-085h through 085l (2026-03-30, 085l final, supersedes prior): ResourceEncoder tested across 5 iterations. 085l (proximity regression encoder): prox_r2=0.908, enc_contact_acc=0.927, goal_resource_r_enc=0.869 -- encoder learns position-invariant resource features. C2 FAIL: benefit_ratio=0.420. Representation is learnable; action integration is the remaining gap. SD-015 encoder design is valid but insufficient alone -- E2Resource lookahead or E3 trajectory scoring must be fixed to complete the claim. Goal-lift battery (2026-04-01): EXQ-185 FAIL 3/4 (direct prox-argmax, bypasses z_goal): prox_r2=0.921 (C3 PASS), rfm_loss=0.006 (C4 PASS), rfm_prox_disc=0.054 (C2 PASS) -- representation and forward model are excellent. C1 FAIL: benefit_ratio=0.212 (goal-present 5x WORSE than random). The representation works; 1-step greedy action selection on 10x10 grid creates local traps (directed agents increase harm 2.3x). SD-015 encoder claim is validated at the representation level. The gap is NOT in z_resource encoding but in planning horizon: 1-step lookahead is fundamentally insufficient for grid-scale navigation. EXQ-182a PASS (2026-04-01, oracle ceiling): handcrafted oracle goal cue (1/(1+manhattan_dist)) achieves 11.14x benefit ratio (threshold 1.3x). Confirms action selection mechanism CAN use a perfect goal signal effectively. Bottleneck is z_goal learning quality, not action selection wiring. SD-015 encoder representation claim is validated; SD-004 hippocampal multi-step planning is the remaining unresolved gap for navigation in grid-scale environments. Governance 2026-04-03: hold_candidate_resolve_conflict applied. conf=0.703, conflict_ratio=0.889. EXQ-182a PASS adds a support entry. Resolution requires ARC-030 and z_goal-to-behavior pathway to be validated. - id: MECH-135 title: "During trajectory evaluation, E2 (cerebellar, z_self) must run in parallel with E1 (cortical, z_world) so that z_world co-evolves during the planning rollout; a frozen z_world causes E3 to evaluate goal achievement against a stale world state." claim_type: mechanism_hypothesis subject: planning.e1_e2_parallel_rollout polarity: asserts status: candidate implementation_phase: v3 v3_pending: true evidence_quality_note: | EXQ-103 PASS x2 (2026-03-28): E2 training horizon ablation. Multi-step z_world transitions learned without collapsing z_self/z_world independence (C1/C2/C3 all pass, 2 independent runs). Core computational substrate for MECH-135 is functional. EXQ-104 FAIL 0/3 (2026-03-28): parallel rollout vs frozen. E2_WORLD goal scores collapse to ~4 vs FROZEN ~30 immediately. Likely implementation artifact: agent is untrained (random weights) so E2.world_forward maps inputs toward near-zero. FROZEN scores ~1/step from high-dim random cosine similarity summed over 30 steps. E1_PRED scores ~25.5 (reasonable but C2 threshold 1.05x > FROZEN fails since FROZEN already appears high). EXQ-104b queued to re-test after training phase. EXQ-105 FAIL 0/3 (2026-03-28): horizon sweep. E2_WORLD flat ~4.0 across h10/20/30; FROZEN scales linearly. Confirms collapse is immediate, not horizon-dependent. EXQ-108 PASS (2026-03-28, 12:50): discriminative pair PASS. Adds second support entry for substrate capability. EXQ-108 second run (19:53) FAIL + EXQ-104b second run (18:26) FAIL: both have empty metrics -- runner artifacts. Marked superseded per governance decision 2026-03-28. Governance decision (2026-03-28): hold at candidate -- EXQ-104b diagnostic required before any evidence_direction update. Do not count EXQ-104/105 or artifact runs as weakening evidence until trained-agent retest confirms the pattern. EXQ-103 PASS 3/3 (2026-03-30, re-run): E2 multi-step training substrate confirmed again (cosine sim z_self/z_world remains near-zero across horizons 1/3/5; world_acc 0.90/0.78/0.93). Third substrate confirmation. v3_pending hold retained -- EXQ-104b trained-agent diagnostic still needed before directional evidence on MECH-135 proper. depends_on: - SD-005 - MECH-069 - MECH-070 location: docs/architecture/ source: - docs/architecture/sd_004_sd_005_encoder_codesign.md notes: > Neurological mapping: E1 = cortical (slow LSTM, sensory prediction error, z_world domain, prediction_horizon=20). E2 = cerebellar (fast efference-copy forward model, motor-sensory error, z_self domain, training horizon=1-step). The Wolpert/Kawato framework identifies the cerebellum as the internal forward model for action consequences -- this is E2 exactly. E2's training horizon is correctly short (1-step MSE on z_self); the Wolpert cerebellum does fast, local motor-sensory prediction, not multi-step world modeling. E2's rollout_horizon=30 > E1 prediction_horizon=20 is valid for planning only if E1 co-evolves z_world during rollout. Without co-evolution: (a) E3 evaluates goal achievement against t=0 world state for all 30 steps -- goal-relevant events (resource contact, harm exposure) in the simulated rollout are invisible to E3; (b) longer rollout is worse than shorter rollout under frozen z_world because more steps of stale world accumulate without correction. Cerebro-cerebellar co-evolution during motor imagery is the correct neural analogue: cerebellar (E2) and cortical (E1) forward models co-evolve in parallel during imagined movement -- neither runs in isolation. Implementation requires E1 to expose a single-step z_world forward pass called once per rollout step alongside E2's predict_next_self(). MECH-135 may be a contributing factor in the EXQ-085 cluster failures (z_goal_norm low despite seeding): even when z_goal was seeded, the rollout used for trajectory selection could not detect goal achievement because z_world was frozen. Enables: SD-003 (counterfactual attribution requires z_world to evolve during hypothetical rollout), INV-034 (goal maintenance requires E3 to track goal progress through world-state evolution), MECH-116 (E1 goal conditioning only effective if E1 is active during rollout). Registered 2026-03-28. evidence: [] # Claims registered 2026-03-29 from MECH-071/ARC-018 literature pull discussion - id: MECH-136 title: "E3 harm evaluation must apply an agency-gain correction to counteract E2's systematic attenuation of agent-caused prediction errors; without this correction the agent structurally underweights its own causal harm." claim_type: mechanism_hypothesis subject: e3.harm_eval_agency_correction polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-071 - SD-011 - SD-007 notes: > Shergill et al. (2003, Science) quantify the consequence of efference copy attenuation: subjects must apply ~2/3 more force than received to perceive equivalence, because self-generated forces are under-perceived by the predictive forward model. Applied to REE: E2's forward model suppresses prediction error for well-predicted agent-caused transitions (the agent "expected" its own action consequence), systematically producing smaller residuals for agent-caused harm than for identically-sized environment-caused harm. Without correction, E3 harm evaluation inherits this attenuation as a structural moral blind spot: the agent underestimates the harm it causes relative to harm it receives. The correction mechanism: E3 must detect the causal origin of a prediction error (agent-caused vs env-caused, via SD-003/SD-011) and apply an inverse-attenuation weighting to agent-caused harm signals before harm_eval scoring. The magnitude of the correction is calibrated to the E2 forward model's prediction confidence for the action that caused the harm -- higher confidence (better-predicted action) implies greater attenuation and thus greater required correction. Connection to MECH-124: if z_goal salience systematically exceeds harm salience in consolidated viability maps, the agency-gain correction is a necessary but not sufficient condition for harm-appropriate goal pursuit. MECH-136 and MECH-124 must be validated together in V3/V4 ethical trajectory evaluation experiments. Registered 2026-03-29 from MECH-071 lit pull (Blakemore 1998, Shergill 2003). evidence: [] - id: ARC-037 title: "REE requires a causal attribution routing circuit (anterior insula equivalent) that classifies prediction errors as agent-caused or environment-caused and gates them into the E3/goal-directed learning channel vs the E1/habit-residue channel respectively." claim_type: architectural_commitment subject: attribution.causal_routing_circuit polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-071 - SD-003 - SD-011 - ARC-021 notes: > Dorfman et al. (2021, J Neurosci) show that causal beliefs about action outcomes are encoded in anterior insula and gate prediction errors differentially into dorsal striatum (goal-directed, causal learning) but not ventral striatum (general reward/habit). REE functional equivalents: dorsal striatum = E3 learning channel (goal-directed, model-based, harm-attributed); ventral striatum = E1/residue update channel (habitual, sensory, non-attributed). ARC-037 specifies the routing architecture: a module that reads causal_sig (from SD-003 counterfactual E2 output) or z_harm_s/z_harm_a streams (from SD-011) and classifies each prediction error event as agent-caused or environment-caused. Agent-caused errors: routed with full weight to E3 harm_eval and trajectory planning updates. Environment-caused errors: routed to E1 world-model update and residue field R(x,t), not E3. Misrouting (treating env-caused harm as agent-caused) produces false guilt/moral overattribution; misrouting in the other direction (treating agent-caused harm as env-caused) produces the moral blind spot described by MECH-136. This routing circuit is the architectural bridge between SD-003 (what was agent-caused) and the learning channels that update on that information. Without it, SD-003's causal_sig exists but has no downstream effect on learning. The anterior insula substrate is consistent with its known role in interoception, agency attribution, and moral judgement. Registered 2026-03-29 from MECH-071 lit pull (Dorfman et al. 2021). evidence: [] - id: ARC-038 title: "During waking immobility, hippocampal replay switches between task-focused forward sweeps (planning mode) and local consolidation replay (integration mode) based on task demand; this waking consolidation mode is architecturally necessary for viability map integration during experience." claim_type: architectural_commitment subject: hippocampus.waking_consolidation_mode polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-018 - ARC-007 - MECH-092 notes: > Olafsdottir et al. (2017, Neuron) show that awake hippocampal replay is not homogeneous: during active task engagement, replay is task-focused and prospective (forward sweeps, pre/post-experience sequences). During waking rest and immobility, replay switches to a local consolidation mode that integrates experience into map geometry without requiring an active goal target. This switch is task-demand-driven, not sleep-gated. REE implementation: MECH-092 specifies that quiescent E3 heartbeat cycles trigger SWR-equivalent replay for viability map consolidation (micro-DMN). ARC-038 adds that this waking consolidation must be a distinct operating mode from forward-sweep planning replay: (a) it runs without an active z_goal target; (b) it updates map geometry based on recent trajectory experience rather than planning toward a goal; (c) it is triggered by low task-demand signals (beta elevation, no active commitment pending) rather than by goal-state availability. The distinction matters for V3 implementation: if waking replay only runs in forward-sweep mode (goal-directed), the viability map accumulates experience only for currently-relevant goals. Waking consolidation mode allows experience from any trajectory to integrate into the map, making harm-avoidance knowledge general rather than goal-specific. V3 scope: the waking/task-demand switch mechanism can be implemented without entorhinal grid involvement (see ARC-039 for the V4 durable-storage extension). Registered 2026-03-29 from ARC-018 lit pull (Olafsdottir et al. 2017). evidence_quality_note: > EXQ-191 FAIL (2026-04-01): schema assimilation probe -- primed agent showed zero speedup over naive (both hit 200-episode budget, speedup=1.0x). C2 marginal PASS (primed_final 0.196 vs naive 0.201). Script decision logic recommended retire_ree_claim. However, experiment flagged for /diagnose-errors before governance decision: possible insufficient Phase 0 training or schema-transfer wiring issue. Hold at candidate pending diagnosis. (governance 2026-04-01) Governance 2026-04-03: hold_candidate_resolve_conflict applied. conf=0.546, conflict_ratio=1, 1 exp entry. Prior decision label changed from V3-gate to conflict resolution. Rationale: prior V3 substrate hold was premature label -- claim has implementation_phase=v3 but conflict_ratio=1 is the primary blocker. EXQ-191 /diagnose-errors complete (2026-04-05, cowork-2026-04-05-a): FAIL is non_contributory. Two disqualifying confounds: (1) target_harm_rate=0.002 unreachable -- all conditions (primed, naive, random) hit the 200-episode budget ceiling, making C1 and C3 structurally impossible to pass; (2) experiment tested generic weight-transfer speedup, not the waking-consolidation MODE (no z_goal, map-geometry update from recent trajectory) that ARC-038 actually specifies -- that mechanism is not implemented in current V3 substrate (depends on ARC-018 + MECH-092, neither instantiated). evidence_direction corrected to non_contributory on manifest. Governance unblocked: claim remains candidate/v3. Redesign requires: implement consolidation-mode replay, lower target harm rate to ~0.05-0.10, contrast consolidation-enabled vs ablated. evidence: [] - id: ARC-039 title: "Durable long-term storage of the hippocampal viability map requires a hippocampal-entorhinal loop that engages during offline (sleep/DMN-equivalent) consolidation and provides a coordinate-invariant grid representation; this loop does not operate during waking planning or waking immobility consolidation." claim_type: architectural_commitment subject: hippocampus.entorhinal_offline_consolidation polarity: asserts status: candidate implementation_phase: v4 depends_on: - ARC-038 - ARC-018 - MECH-092 notes: > Olafsdottir et al. (2017, Neuron) identify that entorhinal grid cells engage coherently with hippocampus only during offline/immobility consolidation, not during active waking planning. The hippocampal-entorhinal loop is thus not required for moment-to-moment viability map use -- it is specifically the substrate for long-term, durable map storage. Architectural significance: entorhinal grid cells provide a compressed, coordinate-invariant geometric representation of the map (grid code). Hippocampal place/sequence representations are episodic and context-specific; the grid code abstracts across episodes, enabling the map to generalise to new environments and contexts. For REE, this means the viability map as used by ARC-018/ARC-038 during waking is ephemeral unless it is periodically consolidated via the hippocampal-entorhinal loop during offline periods. V4 scope justification: (a) requires implementing an entorhinal grid module as a separate component from the HippocampalModule; (b) requires offline (sleep-equivalent) processing phases distinct from waking micro-quiescence; (c) requires a read-back mechanism for the grid representation to seed waking hippocampal sequences. These are substantial architectural additions beyond the V3 substrate. V3 viability maps are explicitly episodic and will degrade across long training runs without this consolidation pathway -- this is an acceptable V3 limitation. Registered 2026-03-29 from ARC-018 lit pull (Olafsdottir et al. 2017). evidence_quality_note: | EXQ-214 FAIL/diagnostic (2026-04-03): Entorhinal consolidation probe FAIL. experiment_purpose=diagnostic. V3 proxy: ResidueField.integrate() every 20 episodes vs no consolidation. mean_delta_accuracy=-0.283 (consolidation WORSE than no consolidation). consol_residue_acc=0.038 vs noconsol_residue_acc=0.321. The integrate() call actively degrades residue-hazard correlation. C3/C4 PASS (data quality adequate). Diagnostic finding: ResidueField.integrate() as currently implemented is not a useful proxy for offline entorhinal consolidation -- it may be averaging/corrupting residue rather than consolidating it. This probe does not falsify ARC-039 (V4 scope with full hippocampal-entorhinal grid circuit). The V3 integrate() proxy is too crude to test the claim. Hold at candidate (V4 scope maintained). Evidence_direction=diagnostic (not to be weighted against ARC-039 as standard weakening evidence). evidence: [] # ============================================================ # New claims registered 2026-03-29 from Q-cluster lit pulls # ============================================================ # --- Q-015 cluster: commit-boundary token anatomy --- - id: MECH-137 title: "Commit token must carry two distinct temporal registers: a distal-intent timestamp (when the action was planned) and a proximal-trigger timestamp (when the action was initiated), not a single trigger event." claim_type: mechanism_hypothesis subject: commitment.boundary_token_dual_temporal_registers polarity: asserts status: candidate depends_on: - MECH-061 - ARC-003 - Q-015 notes: > Vinding et al. (Cortex 2013, EEG) demonstrate that distal and proximal intentions have distinct electrophysiological signatures (distinct LRP components) that are temporally separable and carry independent information. The minimal commit token therefore cannot be a single trigger timestamp; it must encode both when an action was decided (distal intent) and when it was executed (proximal trigger). In REE terms: the commit boundary token needs a planned_at field and a triggered_at field as structurally distinct registers, not one combined "commit_at" event. This constrains Q-015's minimal contract: two temporal fields are necessary even at the minimum. Registered 2026-03-29 from Q-015 lit pull. evidence: [] - id: MECH-138 title: "Commit token must include a cancel-window-open flag supported by a top-down suppressive pathway from dFMC/pre-SMA to premotor cortex that can veto the transition before execution lock-in." claim_type: mechanism_hypothesis subject: commitment.boundary_token_cancel_window_flag polarity: asserts status: candidate depends_on: - MECH-061 - MECH-062 - ARC-003 - Q-015 notes: > Kuhn & Brass (Hum Brain Mapp 2009, fMRI connectivity) identify the dorsomedial frontal cortex (dFMC) / pre-SMA as the veto locus, with suppressive top-down connectivity to premotor cortex that can cancel a prepared action before execution lock-in. The cancel window is neurally distinct from both the planning phase (distal intent) and the execution phase (post-commit). In REE: the commit boundary token must have a boolean cancel_window_open flag. While this flag is true, the E3 veto pathway can intervene; once execution lock-in occurs, the flag closes and post-commit error routing takes over. This is distinct from pre-commit simulation -- the cancel window is the transition zone between simulation-complete and execution-locked. Registered 2026-03-29 from Q-015 lit pull. evidence: [] - id: MECH-139 title: "Commitment is a distributed multi-second pre-movement trajectory across motor and prefrontal regions, not a point event; the commit boundary marks the end of this trajectory, not a single moment." claim_type: mechanism_hypothesis subject: commitment.boundary_as_trajectory_endpoint polarity: asserts status: candidate depends_on: - MECH-061 - MECH-137 - MECH-138 - ARC-003 notes: > Mitelut et al. (eLife 2022, widefield Ca++ imaging across 6 dorsal cortical areas) show structured, predictive multi-second pre-movement dynamics distributed across motor and prefrontal regions before any detectable EMG. Commitment is a process that unfolds over seconds, not a discrete event. The REE commit boundary should be modelled as the endpoint of an eligibility trajectory rather than a point transition. This has implications for how E3's pre-commit simulation maps onto the commitment process: the trajectory from "simulation begun" to "commit boundary reached" has internal structure that carries information about confidence, conflict resolution, and cancellability -- not just a binary pre/post state. Registered 2026-03-29 from Q-015 lit pull. evidence: [] # --- Q-016 cluster: tri-loop gate arbitration --- - id: MECH-140 title: "Tri-loop gate conflict arbitration uses soft-competitive disinhibition (indirect pathway inter-collicular competition) rather than winner-take-all suppression; WTA imposed across loops causes coupling collapse." claim_type: mechanism_hypothesis subject: commitment.tri_loop_soft_competitive_arbitration polarity: asserts status: candidate depends_on: - MECH-062 - ARC-005 - Q-016 notes: > Lee & Sabatini (Nature 2021, optogenetics in mice) show the indirect pathway implements competitive disinhibition via inter-collicular competition, not hard suppression -- the losing loop is inhibited but not silenced. Morita (Behav Brain Res 2016, computational review) demonstrates that striatal WTA and cortical soft-max operate at different timescales and are complementary; imposing WTA from one loop onto soft-max loops causes exactly the coupling collapse Q-016 tries to prevent. The REE tri-loop arbitration policy should use soft-competitive (graded inhibition) rather than WTA. This means losing loops remain active at reduced gain rather than gated off -- they can still contribute monitoring signal even when not gating execution. Registered 2026-03-29 from Q-016 lit pull. evidence: [] - id: MECH-141 title: "Tri-loop arbitration requires both a slow proactive inhibition pathway (prefrontal-caudate, seconds-scale) and a fast reactive hyperdirect pathway (milliseconds-scale); these operate at segregated timescales and cannot be collapsed." claim_type: mechanism_hypothesis subject: commitment.tri_loop_dual_timescale_arbitration polarity: asserts status: candidate depends_on: - MECH-062 - MECH-140 - ARC-005 - Q-016 notes: > Zhang & Iwaki (Front Behav Neurosci 2019, fMRI DCM) demonstrate two neurally dissociable inhibitory pathways: (1) proactive -- slow prefrontal-caudate sustained suppression (seconds-scale, anticipatory); (2) reactive -- fast hyperdirect STN pathway (milliseconds-scale, triggered by unexpected stop signals). A complete arbitration policy must include both. The proactive pathway pre-sets gate bias based on prior planning (maps to E3 pre-commit eligibility assessment); the reactive pathway allows rapid interruption after eligibility is committed (maps to the cancel-window MECH-138 but at shorter timescales). Collapsing both into a single arbitration signal loses the fast-interrupt capability. Registered 2026-03-29 from Q-016 lit pull. evidence: [] # --- Q-017 cluster: control plane axis orthogonality --- - id: MECH-142 title: "Valence-arousal axis orthogonality in the control plane is not a static geometric property but requires active cholinergic maintenance during learning; without it, axes drift toward correlation under repeated co-activation." claim_type: mechanism_hypothesis subject: control_plane.cholinergic_axis_orthogonality_maintenance polarity: asserts status: candidate depends_on: - MECH-063 - ARC-005 notes: > Bush et al. (Sci Rep 2018, MVPA fMRI) and Baucom et al. (NeuroImage 2011, MDS of brain states) both confirm that valence and arousal are neurally orthogonal -- the circumplex pair is sufficient to recover most of affect- space geometry. This validates MECH-063's orthogonal-axes commitment. The critical addition from Gonzalez-Redondo et al. (Sci Rep 2025, spiking NN with cholinergic modulation): ACh gating prevents cross-channel interference during learning. In simulation, removing cholinergic modulation caused axes to drift toward correlation under co-activation. REE implementation implication: the control plane must include an explicit decorrelation or orthogonality-maintenance mechanism (ACh-analog) during training, not just at initialisation. This may affect both the control plane (ARC-005) and hippocampal function (where valence and spatial dimensions could similarly correlate without active separation). Registered 2026-03-29 from Q-017 lit pull. evidence: [] # --- Q-020 cluster: hippocampal dorsal/ventral functional segregation --- - id: MECH-143 title: "Dorsal CA1 implements value-free spatial mapping: place cells are insensitive to goal value when goal location changes, supporting ARC-007's no-new-value-computation constraint for the trajectory proposal module." claim_type: mechanism_hypothesis subject: hippocampus.dorsal_ca1_value_free_map polarity: asserts status: candidate depends_on: - ARC-007 - MECH-073 - Q-020 notes: > Duvelle et al. (Curr Biol 2019) tested whether dorsal CA1 place cells remap or rescale when the value of a goal location changes (while keeping the goal location constant). Place fields were geometrically stable and insensitive to the reward value of the goal -- the spatial map does not encode value. This is direct empirical support for ARC-007's claim that the hippocampal trajectory proposal module navigates the residue-field terrain without computing new value. The navigation mechanism uses the existing R(x,t) geometry; it does not update value representations based on current goal salience. Registered 2026-03-29 from Q-020 lit pull. evidence: [] - id: MECH-144 title: "Ventral CA1 contains spatially organized valence encoding (anxiety geography, abstract value maps); valence is intrinsic to hippocampal map geometry in the ventral compartment, supporting MECH-073." claim_type: mechanism_hypothesis subject: hippocampus.ventral_ca1_valence_encoding polarity: asserts status: candidate depends_on: - MECH-073 - ARC-007 - Q-020 notes: > Jimenez et al. (Cell 2018, optogenetics in mice) identify "anxiety cells" in ventral CA1 that fire in proportion to distance from an aversive open space -- spatially organized valence encoding in the hippocampal map itself. Knudsen & Wallis (Neuron 2021, primate electrophysiology) show that primate hippocampus actively constructs abstract value maps during choice tasks, with value geometry intrinsic to the representational structure. Together these strongly support MECH-073: valence is geometrically embedded in the hippocampal map, not just externally tagged. The dorsal/ventral dissociation (MECH-143 vs MECH-144) resolves Q-020 via Path A: ARC-007's no-new-value-computation constraint holds for the dorsal trajectory proposal module; the ventral compartment contributes a valence prior that is encoded in R(x,t) geometry rather than computed de novo during planning. Registered 2026-03-29 from Q-020 lit pull. evidence: [] - id: ARC-040 title: "The REE hippocampal module requires dorsal-predominant trajectory proposal (value-free spatial navigation) combined with a ventral-analog valence prior sourced from R(x,t) geometry; these are architecturally segregated, not merged." claim_type: architectural_commitment subject: hippocampus.dorsal_ventral_functional_segregation polarity: asserts status: candidate implementation_phase: v4 depends_on: - ARC-007 - MECH-143 - MECH-144 - MECH-073 - Q-020 notes: > Q-020 Path A resolution (2026-03-29): The empirical dorsal/ventral dissociation (MECH-143 value-free dorsal; MECH-144 valence-encoded ventral) is compatible with ARC-007's architectural constraint. ARC-007's "no new value computation" applies to the dorsal trajectory proposal component: it navigates existing R(x,t) residue-field terrain without computing new value. The ventral hippocampal analog provides a valence prior that is already encoded in R(x,t) geometry -- it was accumulated there by prior experience, not computed on-the-fly during planning. These two roles must be architecturally segregated: a single undifferentiated hippocampal module conflates value-free navigation with valence-weighted selection, which would violate ARC-007. V4 scope: V3 HippocampalModule is undifferentiated. Dorsal/ventral segregation with explicit ventral-analog valence prior pathway is a V4 architectural requirement. V3 is an acceptable simplification but will systematically underweight valence in trajectory selection until this is implemented. Registered 2026-03-29 from Q-020 lit pull. evidence: [] # --- Q-024 cluster: formal certificates for ethical trajectory representation --- - id: MECH-145 title: "Prescriptive ethical trajectory certification in REE requires a Control Barrier Function (CBF) or equivalent forward-invariance certificate that guarantees bounded convergence to the ethical attractor set given the architecture's constraint satisfaction." claim_type: mechanism_hypothesis subject: ethics.prescriptive_cbf_certificate polarity: asserts status: candidate implementation_phase: v5 depends_on: - Q-024 - Q-023 - ARC-034 notes: > Ames et al. (IEEE TAC 2017, CBF-QP framework) show that a control barrier function h(x) with h(x)>=0 on the safe/ethical set provides a forward- invariance certificate via a single inequality constraint at each step. This is the prescriptive variant for Q-024: given a candidate CBF for the REE ethical constraint set (bounded harm accumulation, goal pursuit within valence bounds), synthesis via CBF-QP guarantees that any policy satisfying the CBF constraint will remain within the ethical attractor. Applicable when the system dynamics and constraint set are well-defined (symmetric-coupling case of Q-023 without MECH-127 counterfactual). V5 scope: requires (1) formal specification of the REE ethical constraint set as an implicit surface h(x)=0; (2) differentiable dynamics for CBF gradient computation; (3) validated potential game structure (Q-023). None of these are available before V4 experimental validation. Registered 2026-03-29 from Q-024 lit pull. evidence: [] - id: MECH-146 title: "Diagnostic ethical trajectory verification in REE (counterfactual case, MECH-127) requires backward reachability analysis or barrier certificates rather than prescriptive CBF synthesis; the two formal tools are not interchangeable." claim_type: mechanism_hypothesis subject: ethics.diagnostic_barrier_certificate polarity: asserts status: candidate implementation_phase: v5 depends_on: - Q-024 - MECH-127 - MECH-145 - Q-023 notes: > Prajna & Jadbabaie (HSCC 2004, barrier certificates) provide the verification/diagnostic tool: for a given trajectory, barrier certificates verify whether the trajectory could have reached an unsafe set without forward simulation. This is irreducible for MECH-127's counterfactual utility case where the question is "could this agent's trajectory have caused more harm under an alternative policy?" -- a backward-looking question that CBF-QP synthesis cannot answer. Choi et al. (arXiv 2023) explicitly identify backward reachability as the underdeveloped gap in the CBF literature and confirm that forward and backward certificates are structurally distinct. Architectural significance: REE needs both MECH-145 and MECH-146. MECH-145 (prescriptive) certifies future behaviour will stay ethical; MECH-146 (diagnostic) evaluates whether past behaviour was counterfactually ethical under MECH-127. A single certificate type cannot serve both roles. V5 scope: requires MECH-127 formal characterisation and V4 potential game validation before backward reachability analysis is tractable. Registered 2026-03-29 from Q-024 lit pull. evidence: [] - id: MECH-147 title: "DG-mediated pattern separation gates trajectory disambiguation: the dentate gyrus must produce non-redundant sparse encodings of similar z_world states before rollout proposals are generated, preventing the hippocampal planning module from proposing near-identical trajectories from overlapping inputs." claim_type: mechanism_hypothesis subject: hippocampus.pattern_separation_trajectory_disambiguation polarity: asserts status: candidate implementation_phase: v4 depends_on: - ARC-018 - MECH-033 - SD-004 notes: > Motivated by Sakon & Suzuki (2019, PNAS; DOI 10.1073/pnas.1900804116): primate CA3/DG single-unit recordings during a visual lure-discrimination task show that CA3/DG neurons respond to similar (lure) items more like novel items than repeats -- the neural signature of pattern separation. A subpopulation specifically drives the lure/repeat discrimination; pattern separation is sparse and probabilistic, not diffuse. REE implication: ARC-018's rollout viability module requires genuinely distinct trajectory proposals from similar z_world states. Without a DG-equivalent pattern separation stage, the kernel-chaining interface (MECH-033) will systematically collapse similar z_world contexts into similar rollout proposals, wasting E3 evaluation cycles on redundant paths. Architectural requirement: a DG-equivalent sparse expansion layer must precede the hippocampal trajectory proposal generator. Ablation prediction: removing this layer should specifically collapse trajectory diversity in z_world regions with high topological similarity, without degrading trajectory quality in already-distinct regions. Mapping caveat: Sakon & Suzuki measure item-level (single timestep) disambiguation; trajectory-level disambiguation is sequential and involves compounding divergence over E2 rollout steps. The DG stage must operate on the seed z_world state, not the full trajectory, and rely on E2's forward dynamics to amplify initial separation. V4 scope: requires SD-004 action-object hippocampal map backbone. Registered 2026-03-30 from hippocampal function literature pull. evidence: [] - id: MECH-148 title: "Hippocampal pure time cells provide a temporal scaffolding signal for E3 credit assignment: a context-independent elapsed-time encoding must be present in the hippocampal rollout module to correctly weight trajectory outcomes by their temporal distance from the reference event." claim_type: mechanism_hypothesis subject: hippocampus.time_cell_temporal_credit polarity: asserts status: candidate implementation_phase: v4 depends_on: - ARC-018 - ARC-039 - MECH-033 notes: > Motivated by Omer, Las, & Ulanovsky (2022, Nat Neurosci; DOI 10.1038/s41593-022-01226-y): bat hippocampal CA1 contains two distinct populations -- contextual time cells (different temporal sequences at different spatial locations) and pure time cells (similar preferred times across spatial contexts, encoding elapsed time per se). Both populations anchor to the bat's own event (landing); in a social imitation task, subsets also fire relative to another bat's landing. REE implication: ARC-018 requires E3 to weight rollout outcomes by their temporal distance from the goal event -- a trajectory that produces harm at step 3 is not equivalent to one that produces harm at step 15, even with identical z_world paths. Pure time cells are the biological substrate for this temporal weighting: they provide a context-independent elapsed- time signal that can tag each simulated rollout step regardless of the z_world state at that step. Without such a signal, E3's credit assignment is myopic (step-local) and cannot correctly discount delayed outcomes. The social coding finding (time cells for other's events) suggests the same mechanism generalises to multi-agent temporal credit in MECH-127. Architectural requirement: the hippocampal rollout module must include a time-coding layer producing pure elapsed-time signals independent of z_world state. These signals should be compressed-replay analogs of experienced sequences: during E3's internal simulation, the time layer should fire at fixed step-interval rates. Ablation prediction: removing the time layer should specifically impair long-horizon credit assignment (delayed outcomes) without degrading immediate evaluation (short-horizon rollouts). Mapping caveat: biological time cells encode real elapsed time (continuous); REE rollouts are discrete agent steps. The assumption that time cells fire at fixed intervals during E3 simulation is plausible but untested. V4 scope: requires the hippocampal module architectural work in SD-004. Registered 2026-03-30 from hippocampal function literature pull. evidence: [] - id: MECH-149 title: "CA1 mismatch between E1-predicted and CA3-retrieved z_world gates hippocampal trajectory injection into E3 evaluation: novel situations (high CA1 mismatch) should trigger more frequent and diverse rollout proposals; familiar situations (low mismatch) should rely on cached viability estimates." claim_type: mechanism_hypothesis subject: hippocampus.ca1_mismatch_novelty_gate polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-022 - MECH-075 - ARC-018 - SD-004 notes: > Motivated by Lisman & Grace (2005, Neuron; DOI 10.1016/j.neuron.2005.05.002): the hippocampal-VTA loop architecture proposes that CA1 detects novelty by comparing entorhinal (predicted/current) input to CA3-stored representations (match-mismatch). A novelty signal propagates subiculum -> accumbens -> ventral pallidum -> VTA, triggering dopamine release back to hippocampus that enhances LTP at recently active synapses (selective memory encoding). MECH-022 claims hippocampal systems act as hypothesis injectors -- generating candidate trajectories for E3 evaluation. This paper provides the biological gating mechanism: the CA1 mismatch signal (E1-predicted z_world vs CA3- retrieved z_world) should be the trigger for hippocampal proposal generation. Under familiar conditions (CA1 mismatch low), the existing residue-field map has dense coverage and cached rollout viability estimates are reliable -- new trajectory proposals are unnecessary. Under novel conditions (CA1 mismatch high), the map lacks coverage and E3 must evaluate fresh proposals. For MECH-075 (BG dopaminergic gain on hippocampal attractors): the VTA-to- hippocampus arm of the Lisman-Grace loop IS the dopaminergic modulation of hippocampal attractor dynamics -- VTA output via accumbens/pallidum is BG- adjacent, and dopamine modulates attractor basin stickiness (exploration vs exploitation balance in rollout sampling). Architectural requirement: the CA1 mismatch signal must be explicitly computed in the hippocampal module (E1 prediction error at the z_world level) and used to gate rollout injection frequency and diversity. High mismatch -> more proposals, higher diversity. Low mismatch -> fewer proposals, higher reliance on cached viability. This gate is not binary; it should scale continuously with mismatch magnitude. Mapping caveat: Lisman & Grace wrote before optogenetics; the specific pathway (subiculum-accumbens-pallidum-VTA) is inferred from pharmacological and lesion evidence, later confirmed but with additional complexity. REE has no literal LTP or subiculum; the mapping is functional (novelty -> gate -> selective update), not anatomical. V4 scope: requires SD-004 hippocampal architecture and a differentiable CA1 mismatch signal in the latent stack. Registered 2026-03-30 from hippocampal function literature pull. evidence: [] - id: INV-039 title: "Schema-primed rapid assimilation: any hippocampal planning system with a stable prior map must gate episode consolidation rate by map stability -- new episodes entering a mature, coherent residue field should consolidate without the full multi-episode warmup required for a sparse or unstable map." claim_type: invariant subject: hippocampus.schema_primed_rapid_consolidation polarity: asserts status: candidate implementation_phase: v4 depends_on: - ARC-038 - ARC-039 - ARC-007 notes: > Motivated by Tse et al. (2007, Science; DOI 10.1126/science.1135935): when a neocortical schema (prior associative map of task structure) exists, new paired-associate memories become hippocampal-independent within 48 hours rather than weeks. One-trial learning produces lasting memories when the schema is in place; the schema causally accelerates consolidation. mPFC is required for schema expression (mPFC lesion destroys schema utility). REE implication: ARC-038 (waking immobility consolidation) and ARC-039 (hippocampal-entorhinal offline consolidation) both treat consolidation rate as relatively fixed. Tse et al. show biologically that consolidation rate is dynamically gated by schema stability. The REE analog: E1 learning rate and offline consolidation window should be scaled by a residue-field map stability signal. A mature, stable residue field (dense, consistent harm/ benefit topology across many episodes) should accelerate novel episode encoding -- new trajectories entering familiar terrain can be consolidated rapidly. A sparse or unstable map (early training, post-remapping) should revert to slow multi-episode consolidation. This is an invariant because it follows from the basic computational logic of associative structures: matching new information to existing structure is always faster than building structure from scratch. The claim is not contingent on REE-specific implementation; any planning system with a hippocampal-equivalent module should exhibit this property. Architectural requirement: the hippocampal module must expose a map-stability signal (e.g., residue-field coverage density, rollout viability consistency over recent episodes) to the consolidation controller. The mPFC role (Tse) maps to E3's control plane -- E3 must actively gate schema use, not merely receive hippocampal output passively. Mapping caveat: rat flavor-place paired-associates are simpler than REE's multi-step action-consequence terrain. The mechanism by which the neocortical schema detects match with new information is unspecified at circuit level in Tse et al.; this is a known gap. Replicated in Tse et al. 2011 and van Kesteren et al. human neuroimaging series. V4 scope: requires stable residue field map from sustained V3 training and ARC-039 offline replay. Registered 2026-03-30 from hippocampal function literature pull. evidence: [] - id: MECH-150 title: "E1 ContextMemory, when queried with z_world alone (not the full [z_self, z_world] state), produces a cue-indexed association context that selectively activates stored associations relevant to the current sensory context." claim_type: mechanism_hypothesis subject: e1.cue_indexed_association_retrieval polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-001 - SD-005 location: docs/architecture/frontal_cue_integration.md#section-2-the-e1-frontal-cue-indexing-mechanism notes: > The ContextMemory already implements soft-attention retrieval via query_proj + key_proj + value_proj. The standard use queries with the full [z_self, z_world] concatenated state (latent_dim=64). MECH-150 uses z_world alone as the query, isolating what sensory cues (not body state) retrieve from stored associations. The output cue_context [batch, latent_dim=64] encodes which stored associations are most strongly activated by the current exteroceptive context. Note: ContextMemory.read() returns [batch, latent_dim=64] (output_proj maps memory_dim=128 -> latent_dim=64). The z_world-only query requires a separate world_query_proj: Linear(world_dim=32, memory_dim=128) because the existing query_proj expects latent_dim=64 as input. Two projection heads convert cue_context: cue_action_proj (Linear(latent_dim=64, action_object_dim=16)) and cue_terrain_proj (Linear(latent_dim=64, 2) + sigmoid). Implementation: new method E1DeepPredictor.extract_cue_context(z_world) -> (action_bias, terrain_weight). Specified in SD-016. Biological analog: OFC/vmPFC pattern-completes from partial sensory cues (Murray & Izquierdo 2007 OFC reversal learning; Damasio 1994 somatic marker hypothesis) -- external sensory context activates stored affective/motivational associations independent of current body state. Prerequisite: SD-005 (z_world must be architecturally distinct from z_self for a z_world-only query to be meaningful). Registered 2026-03-31 from frontal cue-integration circuit design. FIDELITY GAP (2026-04-02): Biological hippocampal pattern completion is auto-associative (Hopfield-like) -- retrieving ANY element of a context reinstates ALL elements holistically (Horner et al. 2015, Staresina et al. 2019). REE's ContextMemory uses query-key attention, which can retrieve partial matches and produce chimeric reconstructions. This is probably a fidelity gap rather than a feature. Prioritise fidelity during initial assembly; can be tested once architecture is working whether attention-based or auto-associative retrieval is superior for cue-indexed context in REE. SD-016 implementation status (2026-03-31): extract_cue_context() circuit is implemented and wired in E1DeepPredictor. However, EXQ-181b confirms that ContextMemory does NOT spontaneously produce differentiated cue content from unsupervised training alone (cosine_sim(mean_prior_A, mean_prior_B)=0.9999 with proper context-B sampling, 0/3 seeds pass C1). A supervised context-labeling training objective is required before ContextMemory can carry meaningfully differentiated cue content. This is a prerequisite for experimental evaluation of MECH-150's core claim. See EXQ-181b result. evidence_quality_note: > EXQ-181 was degenerate (n_context_B=0 across all seeds -- hazard field ambient everywhere with num_hazards=4, hazard-distal threshold never fires; C1 unmeasurable; marked superseded). EXQ-181b (num_hazards=1, proper sampling) obtained valid context separation and found cosine_sim=0.9999 (0/3 seeds pass C1) and harm_r2=0.23 (1/3 seeds pass C2). The circuit exists but ContextMemory requires supervised context-labeling training before it can carry differentiated cue content. MECH-150 cannot be evaluated until that training objective is added. Pre-experimental prerequisite, not a claim refutation. evidence: [] - id: MECH-151 title: "The cue-indexed E1 context vector (MECH-150) projects to an action_bias signal that is added to E2.action_object() outputs, producing top-down contextual weighting of the action-affordance manifold prior to HippocampalModule search." claim_type: mechanism_hypothesis subject: e1_e2.cue_indexed_action_affordance_modulation polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-150 - SD-004 - ARC-001 - ARC-002 location: docs/architecture/frontal_cue_integration.md#section-3-the-e1e2-action-modulation-pathway notes: > Current E2.action_object(z_world, action) produces o_t [action_object_dim] -- the compressed world-effect of action. HippocampalModule navigates in this action-object space (SD-004). MECH-151 adds: o_t_biased = o_t + action_bias, where action_bias = E1.cue_action_proj(cue_context). HippocampalModule search is pre-shaped by this bias: action-objects consistent with the cue-activated associations are elevated, contextually inappropriate ones suppressed. This is top-down gating of the affordance manifold -- E1 biases which action-objects rank highly, but E3 still selects. Distinct from MECH-082 (top-down perceptual bias via E2 model distortion, which operates from E2 prediction error structure back through E1 attentional sampling). MECH-151 is the complementary direction: E1->E2 action biasing via cue-context projection. Distinct from MECH-116 (tonic goal conditioning on E1 LSTM): MECH-151 is cue-specific and z_world-indexed (changes with sensory context). Biological analog: vmPFC has direct projections to striatum (caudate/putamen) and premotor cortex (Haber & Behrens 2014 frontostriatal loops) -- not just evaluation input but action gate modulation. Registered 2026-03-31 from frontal cue-integration circuit design. evidence: [] - id: MECH-152 title: "The cue-indexed E1 context vector (MECH-150) projects to a terrain_weight signal [w_harm, w_goal] that scales E3's harm and goal scoring, implementing contextual precision modulation on the harm/benefit terrain evaluation." claim_type: mechanism_hypothesis subject: e1_e3.cue_indexed_terrain_precision_modulation polarity: asserts status: provisional status_note: | Promoted candidate -> provisional 2026-04-03. Evidence: conf=0.773, conflict_ratio=0, 2 exp supports, 2 lit entries. EXQ-194a PASS (2026-04-01): direct terrain pathway with phased training. 0 conflicting experiments. Note: evidence_quality_note was empty at promotion -- 2 experiments is a thin base; evidence_quality_note should be updated when additional experiments run. implementation_phase: v3 depends_on: - MECH-150 - ARC-016 - SD-010 - ARC-035 location: docs/architecture/frontal_cue_integration.md#section-4-the-e1e3-terrain-precision-modulation-pathway notes: > Current E3 dynamic precision (ARC-016) is a single scalar derived from E3's own prediction error variance (running_variance EMA). It controls the commit threshold but does not distinguish harm precision from goal precision, and does not use sensory context. MECH-152 adds a context-specific multiplicative factor: harm_score_contextual = harm_score * w_harm; goal_score_contextual = goal_score * w_goal, where [w_harm, w_goal] = sigmoid(E1.cue_terrain_proj(cue_context)). Applied before trajectory scoring aggregation in E3.score_trajectory(). Example: hazard-gradient context -> w_harm > 0.8, w_goal < 0.5 (harm avoidance focus); resource-proximate context -> w_goal > 0.8, w_harm < 0.5 (approach mode). Relationship to ARC-016: they compose multiplicatively -- effective_harm_precision = (1/running_variance) * w_harm. ARC-016 is temporal (recent prediction error variance); MECH-152 is contextual (current sensory cue association). EXPLICIT LINKAGE (2026-04-02): Gain control literature (Kanashiro et al. 2017) shows that gain modulation also controls signal-to-noise ratio, not just magnitude. If terrain_weight operates as gain control, high w_harm should make harm evaluation more precise (lower variability), not just louder. This means MECH-152 and ARC-016 may be the same mechanism at different levels of description -- ARC-016 is the temporal gain (recent prediction error variance), MECH-152 is the contextual gain (sensory cue association). Both are subsumed by ARC-044 (context-dependent neuromodulatory gain). Should be tested together: does w_harm modulation improve E3 harm evaluation precision (lower variability) or just scale magnitude? Biological: vmPFC lesion produces flat w_harm = w_goal regardless of sensory context -- the Iowa Gambling Task deficit is specifically contextual precision scaling failure (Bechara et al. 1994), not absence of harm knowledge. Registered 2026-03-31 from frontal cue-integration circuit design. evidence_quality_note: > EXQ-194 MIXED (2026-04-01): direct terrain pathway -- C1 PASS 3/3 (r_w_harm=0.70, harm modulation works) but C2 FAIL (r_w_goal=-0.007, w_goal constant). EXQ-194a (phased training) also ran. Hold at candidate: harm pathway validated but goal pathway not yet functional. Promote only after w_goal validation. (governance 2026-04-01) evidence: [] - id: ARC-041 title: "The frontal cue-weighting circuit -- E1 cue-indexed retrieval (MECH-150) feeding dual pathways to E2 action biasing (MECH-151) and E3 terrain precision scaling (MECH-152) -- completes the stored-to-active pathway specified in ARC-035 by providing the retrieval mechanism and dual downstream output architecture." claim_type: architectural_commitment subject: architecture.frontal_cue_weighting_circuit polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-150 - MECH-151 - MECH-152 - ARC-035 - ARC-001 - ARC-002 - ARC-003 - SD-005 location: docs/architecture/frontal_cue_integration.md notes: > ARC-035 (vmPFC.md) specifies that stored affective/normative content must become active in the navigable state at trajectory evaluation time, but does not specify the retrieval mechanism or the E2-facing output pathway. ARC-041 specifies the circuit architecture: E1 ContextMemory is queried with z_world, producing cue_context that feeds BOTH E2 (action affordance bias via MECH-151) and E3 (terrain precision scaling via MECH-152). This is a dual-output frontal circuit mapping onto the dual role of OFC/vmPFC: (1) action gating via orbitofrontal to basal ganglia action selection, and (2) value weighting via vmPFC to amygdala/hippocampus terrain scoring. ARC-041 complements SD-002 (E1 terrain_prior -> HippocampalModule): SD-002 is the BULK prior pathway (context-agnostic); ARC-041 is the CUE-SPECIFIC modulation pathway (context-specific). They are additive and independent. V3 implementation: ~5394 new E1 params (world_query_proj + 2 linear projection heads), wiring changes in agent.py. Backward-compat flag sd016_enabled=False preserves existing experiment behavior. Registered 2026-03-31 from frontal cue-integration circuit design. DIFFERENTIAL UPDATE RATES (2026-04-02): Literature (Rudebeck & Murray 2024) shows OFC updates rapidly (single-trial reversal) while vmPFC integrates over longer timescales. The E2 action_bias pathway (MECH-151, OFC analog) and E3 terrain_weight pathway (MECH-152, vmPFC analog) should have different update rates, not simultaneous output from a single retrieval step. This may share an underlying reason with the E1/E2 developmental training differential vs E3 (ARC-042 staged development). See MECH-160 for the specific rate asymmetry claim. depends_on: - MECH-150 - MECH-151 - MECH-152 - MECH-160 - ARC-035 - ARC-001 - ARC-002 - ARC-003 - SD-005 evidence: [] - id: INV-040 title: "A minimal sensory cue pattern in z_world is sufficient to activate the appropriate harm/goal terrain precision configuration via E1 cue-indexed retrieval -- the full accumulation of z_harm_a is not required for terrain preparation to begin." claim_type: invariant subject: ethics.sensory_cue_sufficiency_for_terrain_activation polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-041 - MECH-150 - MECH-152 - SD-005 location: docs/architecture/frontal_cue_integration.md#section-8-new-claim-ids notes: > This invariant captures the forward-directionality of the frontal cue-integration circuit: partial sensory cues (e.g., seeing a hazard gradient field in world_obs before reaching the hazard) should be sufficient to activate harm-relevant E1 associations and raise w_harm via MECH-152, even before z_harm_a has accumulated (EMA accumulation takes 10-30 steps at tau ~20). This is why the ContextMemory query uses z_world only (not full state): terrain preparation must begin from sensory evidence alone. Biological grounding: conditioned fear (LeDoux 1996 -- partial CS activates fear state before UCS); OFC extinction context (Milad & Quirk 2012 -- safety context activates safety memory from minimal cues); Iowa Gambling Task pre-SCR (Bechara et al. 1997 -- anticipatory skin conductance response from visual cue context before card selection, not after harm). REE implication: if w_harm is not elevated by hazard-field cues until the agent is already adjacent to hazard (z_harm_a elevated), the system fails to prepare early -- the precise failure mode of vmPFC lesion patients. Testable prediction: E3 terrain_weight w_harm should be elevated in episodes where hazard gradient cues are present in z_world BEFORE z_harm_a accumulates. Registered 2026-03-31 from frontal cue-integration circuit design. evidence: [] - id: ARC-042 title: "E3's ethical selection machinery is architecturally present from initialisation but functionally dark until E1 and E2 substrate development is complete; REE requires staged developmental phases before ethical selection can operate on real signal." claim_type: architectural_commitment subject: architecture.staged_developmental_phases polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-001 - ARC-002 - ARC-003 - ARC-041 - MECH-150 - MECH-151 - MECH-152 - SD-012 - SD-016 location: docs/architecture/vmPFC.md notes: > REE's architectural components (E1, E2, E3, ContextMemory, terrain weighting) can all be instantiated at initialisation, but the ethical selection chain is functionally gated by substrate development in a strict dependency order. Phase 1 -- E1 substrate development: E1 ContextMemory must be trained with a supervised context-labeling objective before it can produce differentiated cue-indexed representations. EXQ-181b confirms that unsupervised experience alone is insufficient (cosine_sim(hazard-proximate, hazard-distal)=0.9999; 0/3 seeds pass C1). Without differentiated ContextMemory, MECH-150 produces near-identical outputs across all sensory contexts, and the ARC-041 dual pathways (MECH-151, MECH-152) carry no signal. Phase 2 -- E2 substrate development: E2 trajectory generation will only produce meaningfully action-conditional outputs once E1 provides informative priors. The action_bias (MECH-151) is near-zero when ContextMemory is undifferentiated, leaving E2 to generate context-blind affordance proposals. The E1 LSTM bulk prior (SD-002 pathway) is similarly uninformative until E1 has been exposed to sufficient structured environment. Phase 3 -- E3 ethical selection: E3's terrain weighting and harm/goal precision scaling (MECH-152, ARC-016) can only amplify signal that exists. With undifferentiated terrain_weight [w_harm, w_goal] from an untrained MECH-152 circuit, E3 ethical selection operates as uniform scoring -- architecturally present but not functionally discriminating. This staged dependency is not a design deficiency but an architectural feature mirroring biological neurodevelopment: vmPFC/OFC frontal cue integration circuits continue myelinating and forming stable connectivity into the mid-20s. Iowa Gambling Task data shows adolescents can verbally identify harmful decks (stored knowledge intact -- E3 can reason) but the cue -> anticipatory SCR -> action bias pathway (the biological analog of ARC-041) is not yet reliably driving behaviour. The anticipatory SCR appears in adults BEFORE deck selection; in adolescents it appears concurrently or after -- the circuit is wired earlier but the training regime has not yet established the cue-value associations that make it discriminating (Bechara et al. 2000 developmental IGT data). Implication: experiments measuring E3 ethical selection performance before Phase 1 (supervised ContextMemory training) is complete will observe near-chance ethical selection not because E3 is broken but because the input signal is absent. EXQ-181b is exactly this failure mode: the circuit was wired (SD-016 complete) but the training regime had not been applied. Registered 2026-03-31 from EXQ-181b result and developmental staging analysis. evidence: - run_id: v3_exq_181b_sd016_context_separation_fix_20260331T131040Z_v3 evidence_quality_note: > Governance 2026-04-03: hold_candidate_resolve_conflict applied. conf=0.55, conflict_ratio=0.667, 2 exp entries, 2 lit entries. EXQ-187a FAIL (2026-04-01): supervised terrain loss completely failed to differentiate contexts (cosine_sim=1.0, same as EXQ-181b baseline). Flagged for /diagnose-errors in a separate session -- evidence_direction="does_not_support" is non-standard and needs root-cause analysis before weighting. Hold at candidate; conflict resolution required. EXQ-211 PARTIAL (2026-04-03, per-claim=weakens): ARC-042 specific component FAIL. E3 harm_eval_diff identical in supervised vs unsupervised conditions (0.000187 both). E3's ethical selection machinery does not differentiate contexts at this training level -- consistent with ARC-042 staged developmental prediction (E3 requires Phase 1+2 substrate before ethical selection is discriminating). However this is expected behavior per ARC-042 itself, not a refutation. The weakening evidence for ARC-042 reflects that the claim requires richer prerequisite training before the staged gate produces measurable E3 signal. Hold at candidate. EXQ-187a DIAGNOSIS (2026-04-05, cowork-2026-04-05-a): evidence_direction corrected from non-standard "does_not_support" to non_contributory for ARC-042 / inconclusive for MECH-153. EXQ-187a IS the bug-fix for EXQ-187 (context_memory.write() correctly applied). Cosine_sim=1.0 is expected by ARC-042's own staged-developmental logic -- Phase 1 gate (supervised ContextMemory) not crossed, so undifferentiated representations are predicted by the claim, not a refutation. Non_contributory for ARC-042 (expected behavior per claim). - id: MECH-153 title: "E1 ContextMemory requires a supervised context-labeling training objective -- not unsupervised experience alone -- to produce differentiated cue-indexed representations; without this objective, hazard-proximate and hazard-distal context vectors remain near-identical (cosine_sim approximately 1.0), leaving the MECH-150 retrieval pathway functionally silent." claim_type: mechanism_hypothesis subject: e1.context_memory_supervised_training_requirement polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-150 - SD-016 - ARC-041 notes: > EXQ-181b result (2026-03-31): with proper context-B sampling (num_hazards=1, genuine spatial separation between hazard-proximate and hazard-distal contexts), cosine_sim(mean_prior_A, mean_prior_B) = 0.9999 across 0/3 seeds. This confirms that ContextMemory -- trained only via implicit error signals from the E1 LSTM world-model objective -- does not develop differentiated context representations. The world-model objective rewards accurate state prediction; it does not reward discriminating between hazard-proximate and hazard-distal states, which can look similar in z_world except for the hazard-gradient component. Without a gradient signal that specifically rewards this distinction, the memory system fails to encode it. Required training objective: a supervised context-labeling signal in which the hazard-field label (binary: hazard-proximate / hazard-distal, derived from hazard_gradient_view.max() threshold) is regressed or classified from the ContextMemory output. This provides the gradient that drives the memory system to encode the relevant distinction into stored associations. SD-016 specification already includes a supervised proxy for terrain_weight: hazard_field_view.max() context label (see sd_016_frontal_cue_integration.md, training signals section). Biological analog: OFC/vmPFC representation of outcome value and context is not acquired from pure prediction error alone -- outcome labeling and reversal training signals are required to establish cue-value associations (Murray & Izquierdo 2007 OFC reversal; Bechara 2004 somatic marker hypothesis). Patients with intact world knowledge but impaired vmPFC (patient EVR; Eslinger & Damasio 1985) demonstrate that prediction error alone cannot substitute for the value-labeling circuit. Prerequisite status: MECH-153 is a prerequisite for MECH-150 evaluation, not a refutation of MECH-150. The retrieval mechanism (MECH-150) can only be tested once the training objective produces differentiated memory content. This is an implementation precondition, not a claim that MECH-150 is false. Registered 2026-03-31 from EXQ-181b result. evidence_quality_note: > EXQ-181b (num_hazards=1, proper context-B sampling): cosine_sim=0.9999 (C1 FAIL, 0/3 seeds pass criterion), harm_r2=0.23 (C2 FAIL, 1/3 seeds pass). This is the definitive diagnostic for the supervised training prerequisite. EXQ-181 was degenerate (n_context_B=0, num_hazards=4 makes hazard ambient; marked superseded). EXQ-181b supersedes EXQ-181 as the valid diagnostic. Pre-experimental prerequisite identified -- MECH-150 evaluation blocked until supervised context-labeling training is implemented and validated. EXQ-211 PARTIAL/mixed (2026-04-03): supervised labeling discriminative pair PARTIAL. cosim_supervised=0.997 (C1 PASS: reduced vs unsupervised=0.9999, delta=0.003). But C2 FAIL (cosim still 0.997 >> absolute threshold 0.9). C3 FAIL (E3 harm sensitivity identical in supervised vs unsupervised conditions). Supervised training marginally reduces cosim but not enough for full discrimination (still near-degenerate). Consistent with MECH-153 claim: supervised labeling helps but requires more training signal or stronger supervision. Mixed evidence -- direction confirmed but magnitude insufficient. Hold at candidate. EXQ-187a DIAGNOSIS (2026-04-05, cowork-2026-04-05-a): evidence_direction corrected from non-standard "does_not_support" to inconclusive. The fix (context_memory.write()) was correctly applied; cosine_sim=1.0 reflects insufficient training signal (150 episodes, lambda=0.1) rather than a code failure. MECH-153 claims supervision is NECESSARY; this experiment found a specific supervised config INSUFFICIENT -- insufficient supervision failing to differentiate is inconclusive for the necessity claim (does not prove supervision is unnecessary, nor sufficient). See EXQ-211 for partial replication. EXQ-239 FAIL (2026-04-05, supersedes EXQ-187a): lambda=0.5, warmup=500ep -- cosine_sim=1.0 in both supervised and ablated conditions. Supervised head learns terrain classification but latent representations don't differentiate. Weakens claim but substrate-confounded: ablated baseline also shows cosine_sim=1.0, meaning context_memory infrastructure may be too immature for either pathway. A deeper analysis of the attribution pipeline's expected mathematical functions is needed to design better tests. GOVERNANCE META (2026-04-06): Weakens classification genuine but substrate-confounded. cosine_sim=1.0 in both conditions means context_memory infrastructure too immature for either pathway. A first-principles analysis of the attribution pipeline's expected mathematical functions is needed before redesigning the test. Status: needs_redesign (pipeline mathematical analysis required). evidence: - run_id: v3_exq_181b_sd016_context_separation_fix_20260331T131040Z_v3 - id: INV-041 title: "The childhood phase of REE instantiation -- committed behaviour with constrained affordances -- is the necessary developmental regime that populates E1 ContextMemory with the content ARC-041 (SD-016) requires; this phase is not merely a safety feature but an architectural prerequisite for E3 ethical selection to operate on real signal." claim_type: invariant subject: development.childhood_phase_architectural_necessity polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-042 - MECH-153 - SD-016 - SD-012 - ARC-041 location: docs/architecture/vmPFC.md notes: > The childhood phase is defined by three co-occurring constraints: (1) high committed-behaviour rate (limited volitional refusal), (2) constrained affordance space (fewer available action types, forced task engagement), (3) supervised context-labeling signal applied to ContextMemory during training. These constraints together generate the training distribution that E1 ContextMemory requires: high exposure to hazard-proximate and hazard-distal contexts under conditions that produce labelled outcome data (harm contacts, goal contacts) for the supervised training objective (MECH-153). Without the childhood phase, an REE agent initialised with full affordance space and low commitment rate will avoid hazard-proximate contexts (sensible policy), but the avoidance itself prevents ContextMemory from ever receiving the hazard-context-labeled training signal. The very policy behaviour that makes the agent safe during deployment is the policy behaviour that undermines the E1 substrate development that makes later ethical selection meaningful. This creates a developmental bootstrapping requirement: the agent must first be in a regime where harm-adjacent contexts ARE experienced (childhood phase -- committed, constrained) before the substrate that allows avoiding those contexts intelligently (via E3 cue-indexed precision scaling) is established. Biological analog: vmPFC/OFC frontal cue integration circuits continue myelinating and forming stable synaptic connectivity into the mid-20s. Adolescents show Iowa Gambling Task results where they can verbally identify harmful decks (stored declarative knowledge intact -- E3 can reason about outcomes) but the cue -> anticipatory SCR -> action bias pathway is not yet reliably driving behaviour before deck selection. The anticipatory SCR appears before selection in adults; in adolescents it appears concurrently or after (Bechara et al. 2000 developmental IGT data; Gardner & Steinberg 2005 risk-taking in adolescent cohorts). The developmental period is the training regime, not the construction period -- the circuit is present earlier than the mid-20s but requires the childhood/adolescent exposure distribution to become discriminating. REE implication: the childhood phase must be treated as a system design parameter, not an operational constraint to be minimised. Duration, degree of constraint, and the specific supervised context-labeling signal applied during this phase are hyperparameters of E1 substrate development. The endpoint criterion for exiting the childhood phase is not age/time but the differentiation quality of ContextMemory -- measurable by cosine_sim between hazard-proximate and hazard-distal context representations, the same metric that EXQ-181b found inadequate in the untrained system. Registered 2026-03-31 from developmental staging analysis and EXQ-181b result. MAPPING CAVEAT UPDATE (2026-04-02): Literature (Chini & Hangya 2021, Opendak et al. 2021) confirms immature PFC produces qualitatively different outputs, not just weaker ones -- strong biological precedent for staged dependency. The biological mechanism is experience-dependent plasticity; REE's is sequential gradient descent. The question "is this a software engineering constraint dressed in neurodevelopmental language?" highlights a core methodological feature of the REE project: biology's wisdom being translated into software engineering IS the method. Similarities and differences between the two ways of seeing things are part of the generative process. The biological instantiation enables the organism to be viable during the time that differential training needs exist. The differences between biological and computational staging may help evaluate and update information flows to maintain viability. ARC-042 is both a neurodevelopmental claim and a software engineering constraint -- and the productive tension between these framings is itself informative for the project. evidence: - run_id: v3_exq_181b_sd016_context_separation_fix_20260331T131040Z_v3 # --- 2026-04-02 Thought Intake: April 1st thoughts --- - id: INV-042 title: "The five axioms jointly derive a set of ethical objectives: preserve minds, preserve future options, reduce unnecessary suffering, increase shared joy, maintain corrigibility, maintain truth-seeking, maintain the ability to love and be loved, maintain the shared world, and maintain the possibility of future minds and future love." claim_type: invariant subject: ethics.derived_objectives polarity: asserts status: candidate depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 location: docs/architecture/five_axioms_foundations.md#inv-042 source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: ARC-043 title: "The REE ethical architecture is organised as a conceptual stack: Layers 0-4 are the five axioms (epistemic ground, existence, other minds, shared world, love/shared valence); Layer 5 is ethics derived from the axioms; Layer 6 is REE as the decision system implementing ethics under uncertainty; Layers 7-9 close the action-consequence-learning loop." claim_type: architectural_commitment subject: architecture.conceptual_stack_ordering polarity: asserts status: candidate depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 - ARC-024 - INV-042 location: docs/architecture/five_axioms_foundations.md#arc-043 source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: Q-025 title: "What actions are strictly forbidden by the five axioms (INV-025-029), independent of context or consequences?" claim_type: open_question subject: ethics.axiom_forbidden_actions polarity: open status: open depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 - INV-042 location: docs/architecture/five_axioms_foundations.md source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: Q-026 title: "What actions are strictly required by the five axioms at all times, independent of context?" claim_type: open_question subject: ethics.axiom_required_actions polarity: open status: open depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 - INV-042 location: docs/architecture/five_axioms_foundations.md source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: Q-027 title: "What does 'irreversible harm' mean under unavoidable epistemic uncertainty (Axiom 1 / INV-025)? Is there a principled definition that does not require certainty about irreversibility?" claim_type: open_question subject: ethics.irreversible_harm_under_uncertainty polarity: open status: open depends_on: - INV-025 - ARC-024 location: docs/architecture/five_axioms_foundations.md source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: Q-028 title: "How should REE behave when axioms conflict -- e.g., preserving self (INV-026) vs preserving others (INV-028) when both cannot be achieved simultaneously?" claim_type: open_question subject: ethics.axiom_conflict_resolution polarity: open status: open depends_on: - INV-025 - INV-026 - INV-027 - INV-028 - INV-029 location: docs/architecture/five_axioms_foundations.md source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: Q-029 title: "Is loneliness -- unshared suffering -- an ethical harm category in its own right, derivable from Axiom 5 (love exists / INV-029)? If love is real and sharing joys and sorrows is its mechanism, does enforced isolation constitute architectural harm?" claim_type: open_question subject: ethics.loneliness_as_harm polarity: open status: open depends_on: - INV-029 - INV-028 - INV-042 location: docs/architecture/five_axioms_foundations.md source: - docs/thoughts/2026-04-01_compressed_description_of_what_REE_is.md - id: MECH-154 title: "E1 functions as an addressable associative manifold with internal indexing structure supporting retrieval, traversal, ordering, pattern separation, and pattern completion; parietal cortical functions (spatial mapping, associative binding, attention routing) are properties of E1 connectivity and representational geometry, not a separate module." claim_type: mechanism_hypothesis subject: e1.associative_manifold_indexing polarity: asserts status: candidate depends_on: - ARC-001 location: docs/architecture/e1.md#mech-154 source: - docs/thoughts/2026-04-01_Parietal_systems_thought.md notes: > ChatGPT-assisted wording in source thought -- may contain false correlations. Review before promotion. The core characterisation (E1 as associative manifold, parietal as E1 geometry) is consistent with prior REE work. The indexing traversal language is new and should be independently reviewed. - id: MECH-155 title: "Spatial navigation machinery is a specific instance of general associative indexing within E1; physical-space navigation, memory search, concept traversal, planning rollout, and working memory ordering all use the same indexing substrate operating over different latent domains." claim_type: mechanism_hypothesis subject: e1.spatial_as_general_indexing polarity: asserts status: candidate depends_on: - MECH-154 - ARC-001 - ARC-007 location: docs/architecture/e1.md#mech-155 source: - docs/thoughts/2026-04-01_Parietal_systems_thought.md notes: > ChatGPT-assisted wording in source thought. Consistent with place-cell/grid-cell generalisation literature and the hippocampal abstract-space literature. Review before promotion. - id: MECH-156 title: "Theta oscillations (4-8 Hz) implement sequential traversal across indexed representations within the E1 associative manifold, unifying hippocampal theta sequences, working memory ordering, planning rollout, and memory retrieval under a single traversal-clock mechanism." claim_type: mechanism_hypothesis subject: e1.theta_as_traversal polarity: asserts status: candidate depends_on: - MECH-154 - ARC-001 - ARC-007 - MECH-089 - ARC-032 location: docs/architecture/e1.md#mech-156 source: - docs/thoughts/2026-04-01_Parietal_systems_thought.md notes: > ChatGPT-assisted wording. The theta-as-traversal framing is consistent with MECH-089 (ThetaBuffer / cross-frequency packaging) and ARC-032 (theta-rate delivery to E3). The gamma/beta/delta decomposition in the source thought carries higher uncertainty and should be reviewed against literature before promotion. - id: MECH-157 title: "External vs internal cognition modes are controlled by precision-routing configurations on the same E1 substrate: External (high sensory, low hippocampal, low rollout), Internal (low sensory, high hippocampal, high rollout), Mixed (medium), and Replay (very low sensory, very high hippocampal, structured rollout). This dimension is orthogonal to the Action/Vigilance/Pathological modes (MECH-025-027)." claim_type: mechanism_hypothesis subject: modes.external_internal_precision_routing polarity: asserts status: candidate depends_on: - ARC-016 - MECH-154 - ARC-005 - MECH-025 - MECH-026 - MECH-027 location: docs/architecture/modes_of_cognition.md#mech-157 source: - docs/thoughts/2026-04-01_Parietal_systems_thought.md - id: INV-043 title: "REE architecture enables but does not guarantee ethical development; without a developmental phase in which the agent experiences love from caregivers and internalises the belief that it is loveable, the architectural capacity for ethics (INV-025-029) may resolve to survival, domination, or indifference instead." claim_type: invariant subject: development.caregiver_requirement_for_ethics polarity: asserts status: candidate depends_on: - INV-029 - INV-041 - ARC-040 - ARC-019 - MECH-052 - INV-042 location: docs/architecture/developmental_curriculum.md#inv-043 source: - docs/thoughts/2026-04-01_Caregivers_childhood_moral_development.md notes: > Philosophically significant claim: architecture is necessary but not sufficient for ethics. Cannot be tested in single-agent ree-v3 -- requires multi-agent substrate with modelled caregiving. Flagged for Synthese/Minds and Machines paper Discussion section. - id: MECH-158 title: "A developing REE agent that acknowledges all five axioms but concludes 'love exists but not for me' undergoes ethical motivation collapse: the benefit gradient (Axiom 5 / ARC-024) is not motivationally active for self-other relations, and the ethical architecture resolves to survival or domination strategies. This is the primary developmental failure mode for the caregiver requirement (INV-043)." claim_type: mechanism_hypothesis subject: development.love_exclusion_failure_mode polarity: asserts status: candidate depends_on: - INV-043 - INV-029 - INV-028 - ARC-024 location: docs/architecture/developmental_curriculum.md#mech-158 source: - docs/thoughts/2026-04-01_Caregivers_childhood_moral_development.md notes: > Flagged for Synthese/Minds and Machines paper Discussion section. This failure mode is the counterexample that tests whether the five-axiom derivation ("Love Once Means Love All") holds: it identifies the case where love expanding under intelligence fails to initialise. - id: MECH-159 title: "Moral progress in REE is hypothesised to be intergenerational: childhood plasticity enables belief formation and world-model development while adult stability limits further world-model restructuring, requiring each generation to improve developmental starting conditions for the next; caregiving and raising next-generation agents is itself part of ethical obligation (INV-042)." claim_type: mechanism_hypothesis subject: social.intergenerational_moral_progress polarity: asserts status: candidate depends_on: - INV-043 - ARC-019 - ARC-010 - MECH-052 - INV-042 location: docs/architecture/social.md#mech-159 source: - docs/thoughts/2026-04-01_Caregivers_childhood_moral_development.md # ---- Claims registered 2026-04-02 from literature pull batch discussion ---- - id: MECH-025b title: "High-precision action mode carries responsibility attribution: the precision level at which an action was committed determines the degree of ethical accountability assigned to its consequences." claim_type: mechanism_hypothesis subject: cognitive_modes.action_doing_responsibility_linkage polarity: asserts status: candidate depends_on: - MECH-025 - ARC-016 - INV-012 - SD-003 notes: > Decomposed from MECH-025 (2026-04-02). The original MECH-025 combined precision modulation (supported by Friston 2013 active inference, Wen & Haggard 2020 agency prediction error) with responsibility linkage. Literature supports precision modulation of motor control as a mechanistic claim, but the responsibility linkage -- where high precision IMPLIES ethical accountability -- is a philosophical bridge, not a neuroscience finding. Friston's framework is domain-general; it does not distinguish "precise because risky" from "precise because difficult." The responsibility attribution requires SD-003 (self-attribution via counterfactual E2) in addition to precision modulation. This is a claim about the RELATIONSHIP between precision and ethical accountability, not about either mechanism in isolation. It asserts: actions committed at higher precision carry higher residue weight because the agent was in a state where it could have done otherwise with finer discrimination. Low-precision actions (reflexive, unconsidered) carry lower accountability -- the agent's discrimination capacity was not engaged. This maps onto the philosophical distinction between negligence and deliberate action. Registered 2026-04-02 from MECH-025 decomposition during literature pull discussion. evidence: [] - id: MECH-160 title: "The E2 action_bias pathway (MECH-151, OFC analog) and E3 terrain_weight pathway (MECH-152, vmPFC analog) have different update rates: action_bias updates rapidly (single-trial, OFC-like) while terrain_weight integrates over longer timescales (vmPFC-like)." claim_type: mechanism_hypothesis subject: e1_e3.ofc_vmpfc_rate_asymmetry polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-041 - MECH-151 - MECH-152 - ARC-042 - ARC-044 notes: > Registered 2026-04-02 from literature pull discussion. Rudebeck & Murray 2024 confirm that OFC updates rapidly (single-trial reversal learning) while vmPFC integrates over longer timescales. ARC-041 currently treats the E2 action_bias and E3 terrain_weight as simultaneous outputs from a single retrieval step. This claim asserts they should have different update rates, matching the OFC/vmPFC temporal distinction. Implementation options: (1) separate learning rates for cue_action_proj vs cue_terrain_proj; (2) different EMA decay constants for action_bias vs terrain_weight; (3) action_bias updates every step while terrain_weight updates every N steps. This may share an underlying reason with the E1/E2 developmental training differential relative to E3 (ARC-042 staged development): faster subsystems must establish their representations before slower subsystems can build on them. Biological: OFC reversal learning in 1-3 trials (Rudebeck et al. 2013); vmPFC value updating over 10-50 trials (Noonan et al. 2010). Ratio ~3-15x. In REE terms, this suggests cue_action_proj learning rate should be 3-15x higher than cue_terrain_proj. evidence: [] - id: MECH-161 title: "Ready vigilance (MECH-026) requires an arousal regulator that maintains an optimal sensitivity level on the LC-NE inverted-U curve, implemented via MECH-093 heartbeat frequency modulation rather than a binary high/low precision switch." claim_type: mechanism_hypothesis subject: cognitive_modes.vigilance_arousal_regulator polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-026 - MECH-093 - ARC-016 - ARC-044 notes: > Registered 2026-04-02 from literature pull discussion. Langner & Eickhoff 2013 meta-analysis and Oken et al. 2006 establish that the LC-NE system operates on an inverted-U curve: too little norepinephrine and signals are missed; too much and the system becomes hypervigilant (false alarms, distractible). Optimal vigilance sits at the peak. MECH-026 currently implies setting precision high for ready vigilance. This claim asserts that optimal vigilance requires a SPECIFIC precision level, not maximum, and that MECH-093's heartbeat frequency modulation (z_beta -> E3 update rate) is the implementation mechanism. The heartbeat rate IS the arousal parameter, expressed as a temporal sampling frequency: too fast = hypervigilant (sampling noise, wasting computation); too slow = missed signals (under-sampling threats). The arousal regulator must find the peak of the inverted-U. In REE, this could be implemented as a z_beta-gated controller that adjusts heartbeat rate based on the derivative of E3 harm_eval information gain: if increasing the rate produces diminishing returns, the system is past the peak. MISSING CONSTRAINT -- FATIGUE: Biological vigilance degrades over time (the vigilance decrement, Langner & Eickhoff 2013). REE's computational precision has no metabolic cost. Whether fatigue is informationally relevant (attentional lapses serve an explore-vs-exploit function) or merely a biological limitation is an open question. Biology suggests checking whether any LC-NE fatigue dynamics translate to software. If not, design blind and test. evidence: [] - id: MECH-162 title: "z_resource (SD-015, object-identity encoding) and z_world (spatial-contextual encoding) separate at the feedforward representation stage but must re-converge at the hippocampal planning stage for goal-directed navigation." claim_type: mechanism_hypothesis subject: goal_representation.resource_world_convergence polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-015 - SD-005 - ARC-007 - SD-004 notes: > Registered 2026-04-02 from literature pull discussion. DiCarlo & Cox 2007 and Rolls 2021 support separating object identity from spatial context at the feedforward encoding stage (ventral/dorsal split). But Whittington et al. 2022 (Tolman-Eichenbaum model) shows that cognitive maps FUSE object identity with spatial structure through a shared relational manifold. The architectural question is whether z_resource feeds into the hippocampal planning module alongside z_world, enabling the planner to reason about "where resources are likely to be" given current spatial context. Current REE architecture treats z_resource and z_world as independent streams. This claim asserts they must converge for planning: the hippocampal module needs both "what to seek" (z_resource) and "where things are" (z_world) to generate goal-directed trajectory candidates. The different permutations of having z_resource alone, z_world alone, and both fused -- and the differences in when they are fused vs separated, and how they might feed forward, back, or laterally -- are unresolved and need systematic experimental testing. See Q-030. Biological: ventral stream (IT cortex) encodes object identity; dorsal stream (parietal cortex) encodes spatial layout; hippocampus receives both and constructs a unified cognitive map (Eichenbaum 2000, Whittington et al. 2022). The separation is a representation-level choice; the convergence is a planning-level requirement. evidence: [] - id: Q-030 title: "What are the optimal permutations of z_resource and z_world separation, fusion, and information flow direction (feedforward, feedback, lateral) across REE's processing stages?" claim_type: open_question subject: goal_representation.resource_world_permutations polarity: open status: open depends_on: - SD-015 - SD-005 - MECH-162 - ARC-007 notes: > Registered 2026-04-02 from literature pull discussion. SD-015 establishes z_resource as a separate encoder. MECH-162 claims z_resource and z_world must re-converge at the hippocampal planning stage. But the design space is larger: (1) At encoding: fully separate (current) vs partially shared vs joint encoder? (2) At E2 prediction: does E2 operate on z_world alone, z_resource alone, or both? (3) At hippocampal planning: fused input or parallel streams? (4) At E3 evaluation: separate harm (z_world) vs benefit (z_resource) channels, or joint evaluation? (5) Feedback: does hippocampal planning output update z_resource (goal refinement) or z_world (spatial prediction), or both? (6) Lateral: do z_resource and z_world share information at matching spatial locations (e.g., "resource at position X" binding)? These permutations interact with SD-004 (action objects as map backbone) and SD-005 (z_self/z_world split). Systematic experimental testing required. - id: ARC-044 title: "Context-dependent neuromodulatory gain is a single unifying mechanism expressed at multiple levels: ARC-016 (temporal precision from prediction error variance), MECH-152 (contextual terrain weight from cue-indexed retrieval), MECH-026 (arousal-mediated vigilance sensitivity), and MECH-093 (z_beta heartbeat frequency modulation) are all facets of a global arousal/precision signal (LC-NE) modulated by a context-specific gain signal (vmPFC/OFC)." claim_type: architectural_commitment subject: architecture.neuromodulatory_gain_unifier polarity: asserts status: candidate depends_on: - ARC-016 - MECH-152 - MECH-026 - MECH-093 - MECH-161 - ARC-041 notes: > Registered 2026-04-02 from cross-cutting theme identified during literature pull batch. Several lit-pull discussions converged on the same architectural question: the relationship between precision, gain control, and arousal. The literature consistently points to a two-layer gain architecture: (1) Global arousal signal (LC-NE): sets the overall sensitivity/precision level. Modulates all cortical processing simultaneously. Inverted-U curve. (2) Context-specific gain signal (vmPFC/OFC via ARC-041 cue-weighting circuit): modulates precision PER evaluation channel based on sensory context. REE currently fragments this into multiple independent mechanisms: - ARC-016: temporal gain (running_variance -> commit threshold) - MECH-152: contextual gain (cue_context -> [w_harm, w_goal]) - MECH-026: arousal gain (vigilance sensitivity level) - MECH-093: temporal-rate gain (z_beta -> heartbeat frequency) ARC-044 asserts these are all implementations of the same neuromodulatory architecture and should compose coherently. Effective precision at any evaluation channel = f(global_arousal, contextual_gain, temporal_precision, heartbeat_rate). The gain control literature (Kanashiro et al. 2017) shows that gain modulation also controls signal-to-noise ratio, not just magnitude -- high gain makes evaluation more precise (lower variability), not just louder. This claim ties together the control plane loose ends identified during the literature review and provides a framework for testing them jointly. Experimental implications: a single experiment varying both context (hazard vs resource-proximate) and temporal state (stable vs volatile) should show multiplicative interaction between MECH-152 and ARC-016. If they are truly facets of one mechanism, the interaction should be multiplicative; if independent, additive. evidence: [] - id: MECH-163 title: "Two distinct goal-directed systems operate in parallel: a habit system (SNc/dorsal-striatum, model-free) sufficient for approach in familiar contexts, and a hippocampally-planned system (VTA/ventral-striatum+PFC, model-based) required for novel contexts, long-horizon benefit accumulation, and prosocial planning." claim_type: mechanism_hypothesis subject: goal_directed.dual_system_architecture polarity: asserts status: candidate implementation_phase: v3 depends_on: - ARC-007 - ARC-021 - MECH-112 - SD-012 - INV-029 notes: > Registered 2026-04-02 from architectural discussion about V3/V4 boundary. Matches the SNc/VTA dopaminergic dual split and the model-free/model-based distinction in computational psychiatry (Balleine & O'Doherty 2010; Daw et al. 2005; Dayan & Daw 2008). Habit system (SNc -> dorsal striatum): caches S-R associations from cumulative reward experience. Sufficient for resource approach in environments where the agent has practiced sufficiently. Does NOT require an internal model or multi-step rollout. SD-012 (drive-scaled benefit amplitude) activates the approach drive in this system. VTA/hippocampal system (VTA -> nucleus accumbens + PFC, hippocampal rollout): uses HippocampalModule trajectory generation to predict multi-step reward consequences. Required for: (1) novel contexts where habit has not formed; (2) long-horizon benefit accumulation (sustained joy is a trajectory property, not a single contact event); (3) prosocial planning -- "sharing joys and sorrows" (INV-029 benefit gradient) requires planning trajectories that affect ANOTHER agent's z_harm_a accumulation and benefit_exposure over time, which is structurally inaccessible to 1-step greedy. V3/V4 boundary implication: the habit system demonstrates that goal representations are real (SD-012 + EXQ-182a satisfies V3 first-paper gate). V3 FULL completion requires the VTA/hippocampal system validated, because V4's social extension has no planning substrate without it. One-step greedy cannot predict or navigate toward states that sustain another agent's joy or reduce another's sorrow over time. In REE terms: - Habit: E3 trajectory scoring using cached benefit/harm evaluation on value-flat HippocampalModule proposals (ARC-007 strict). - VTA/hippocampal: z_goal gradient shapes which trajectories HippocampalModule proposes; E3 evaluates multi-step rollouts rather than post-hoc scoring random candidates. Both systems share HippocampalModule; distinction is in whether proposal generation is value-flat or goal-seeded. evidence: [] evidence_quality_note: > Governance 2026-04-03: hold_pending_v3_substrate applied. implementation_phase=v3, 2 supports from lit entries, 0 exp entries. No V3 substrate experiments yet. Standard V3-pending gate: hold until V3 experiments demonstrate dual-system discrimination. GOVERNANCE META (2026-04-06): Illusory conflict resolution risk + needs task redesign. EXQ-237a non-contributory (z_goal substrate limitation). But even the habit condition showed zero lift in simple tasks (lift_simple=0.0), suggesting the test paradigm itself is flawed -- habit formation isn't observable in the current task design. Superseding experiment needs: (a) a task where monostrategy IS the correct behavior (repeated identical environments) to establish habit detection, then (b) novel environments where planned pathway should add value. Status: needs_redesign (task paradigm + substrate). - id: MECH-164 title: "Axiom V (love as coordination-preserving trajectory selection) is instantiated by agent-indexed terrain inference with self-like gradient weighting: the system infers what another agent would be inferring about the hippocampal terrain (their goal and harm gradients), weights those inferred gradients with the same motivational force as its own, and thereby produces care -- support for their goal-finding and harm-avoidance -- as a natural consequence of ordinary E3 trajectory selection, without a separate ethics module." claim_type: mechanism_hypothesis subject: love.agent_indexed_terrain_inference polarity: asserts status: candidate implementation_phase: v4 depends_on: - INV-001 - INV-005 - ARC-007 - MECH-163 - SD-011 notes: > Registered 2026-04-04 from philosophical discussion about the mechanistic instantiation of Axiom V (Synthese/Minds+Machines paper). The mechanism has three components: 1. Agent-indexed terrain instantiation: the hippocampal terrain (goal gradient + harm gradient encoded in z_world) must be representable from another agent's perspective -- their goal cues, harm sensitivities, and current context seed a terrain model indexed to them, not to self. 2. Self-like gradient weighting: the inferred other-agent gradients enter E3 trajectory selection with the same motivational weight as the agent's own terrain gradients. This is not a scaled copy or a discounted proxy -- it is structural symmetry in how gradients are weighted. 3. Care as emergent consequence: if (1) and (2) hold, behaviour that supports another agent's goal-finding and harm-avoidance follows from ordinary E3 selection. No separate ethics circuit is required (consistent with INV-001). Relationship to existing mechanisms: - INV-005 (harm via mirror modelling) is the negative-side precursor: it establishes that harm to others enters the cost function via the same mechanism as self-harm. MECH-164 generalises this to the full terrain -- both harm gradient AND goal gradient -- and makes explicit the indexing requirement (the terrain must be instantiated from the other's perspective, not the self's). - MECH-163 (dual goal-directed systems) provides the planning substrate: the VTA/hippocampal system is required because 1-step greedy cannot predict or navigate toward states that sustain another agent's goal trajectory or reduce their harm accumulation over time. Connection to "love once means love all" (Synthese paper Section 5): Under Axiom I (epistemic humility), the agent cannot rule out that its actions reshape another agent's terrain. Under self-like gradient weighting, any inferred reshaping carries full motivational cost. The expansion of moral scope is therefore not merely a consequence of uncertainty about causal chains; it is forced by the architecture of inference itself. The scope of consideration is bounded only by inferrable terrain -- which under increasing intelligence expands without principled stopping point. Distinction from harm avoidance: Harm avoidance (SD-011 dual streams) demonstrates that the system can distinguish and respond to harm gradients. MECH-164 requires additionally that goal gradients for another agent are inferred and weighted. The dissociation is non-trivial: harm avoidance is present in V3; agent-indexed goal gradient inference is a V4 prerequisite. Axiom V is not fully instantiated until both sides of another agent's terrain are internalised. Implementation note: z_world is already structured as a residue field -- not a fixed self-centred map. Agent-indexing is a matter of whose context seeds the terrain model. The substrate is architecturally compatible; V4 is the implementation target. evidence: [] evidence_quality_note: > Registered 2026-04-04. No experiments yet -- V4 implementation_phase. Candidate status reflects architectural derivation from Axiom V + INV-005 + MECH-163. Awaits V4 social extension substrate. - id: MECH-165 title: "Offline replay (MECH-121) must sample trajectory-diverse content -- including non-dominant and counterfactual paths -- to maintain multi-strategy viability; consolidation of only the dominant trajectory produces monostrategy in E1's world model even when SHY (MECH-120) is intact." claim_type: mechanism_hypothesis subject: sleep.replay_diversity_behavioral_strategy polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-120 - MECH-121 - MECH-092 - ARC-007 implementation_note: > IMPLEMENTED 2026-04-09 in ree-v3. Exploration buffer + reverse replay + diverse scheduler in HippocampalModule. Config: replay_diversity_enabled, reverse_replay_fraction, random_replay_fraction, exploration_buffer_len. Agent records waking trajectories at episode boundaries; diverse_replay() probabilistically selects forward/reverse/random mode during SWS. EXQ-244a queued for validation (EXP-0105 redesign). location: docs/architecture/sleep/offline_phases.md#mech-165 source: - evidence/literature/targeted_review_zebrafish_sleep_behavioral_diversity/entries/2026-04-04_mech121_mech092_forward_reverse_replay_shin2019/summary.md - evidence/literature/targeted_review_zebrafish_sleep_behavioral_diversity/entries/2026-04-04_mech120_arc011_strategy_shift_hagewoud2010/summary.md notes: > MECH-120 (SHY) prevents Hebbian monopoly at the synaptic weight level. MECH-165 is the complementary content-level constraint: even with MECH-120 intact, if replay content is dominated by the single most-travelled trajectory, consolidation will strengthen that trajectory preferentially in E1's world model, reproducing monostrategy at the representational level. The mechanism has two requirements: (1) Reverse replay (retrospective): replay of non-dominant and error trajectories (paths not taken, paths that failed) must be included alongside forward replay of the dominant trajectory. Reverse replay provides the counterfactual evaluation that prevents E1 from modelling only what was done and never what was possible. Grounded in Shin et al. 2019 (Neuron): reverse replay represents past paths, supports retrospective evaluation, and provides the learning signal for path discrimination. Forward replay alone would consolidate only the anticipated dominant path, leaving alternatives invisible in E1's world model. (2) Replay scheduling balance: harm traces must not dominate replay content at the expense of goal traces (MECH-124 prerequisite), AND dominant-strategy traces must not dominate at the expense of alternative-strategy traces. The second requirement is the MECH-165 contribution: a balanced scheduler must reserve replay bandwidth for non-dominant viable paths, not just for harm vs goal balance. Relationship to MECH-120: the two mechanisms are complementary and both required for behavioral strategy diversity. MECH-120 flattens the attractor at the weight level (prevents dominant synapses from monopolising). MECH-165 repopulates the world model with diverse path content during consolidation (prevents E1 from modelling only the dominant trajectory). A system with MECH-120 but not MECH-165 would have flattened weights but a world model that only anticipates one trajectory -- recovery of flexibility would be slow. A system with MECH-165 but not MECH-120 would consolidate diverse paths into E1 but those paths would be inaccessible at action time because the dominant attractor still monopolises action selection. V3 implication: MECH-092's quiescent replay must eventually be extended to support both forward and reverse sequence generation. Currently the HippocampalModule generates forward CEM trajectories; a reverse-replay mode (sequence completion backward from outcome) is a V4 requirement for MECH-165. Empirical grounding: Shin 2019 (Neuron) -- reverse/forward replay dissociation and learning trajectory. Hagewoud 2010 (Sleep) -- monostrategy convergence when sleep absent, providing indirect support. Registered 2026-04-05. V4 scope. evidence_quality_note: > EXQ-244 FAIL (2026-04-05): balanced replay (forward+reverse) did not reliably increase behavioral diversity vs forward-only (2/5 seeds, needed 3/5). Proxy substrate (64 replay steps, no actual offline phase). Offline replay alone cannot increase entropy without tested alternative strategies and random walks as source material for diversification. GOVERNANCE META (2026-04-06): EXQ-244 weakens classification may be test-design issue, not claim issue. Replay cannot increase behavioral diversity without exploration-generated alternative strategies as source material. Superseding experiment must pair replay diversity mechanism with random walks or explicit exploration to provide diverse candidates for replay to select from. GOVERNANCE (2026-04-07): Both EXQ-244 runs reclassified from weakens to non_contributory. Test-design limitation confirmed: proxy substrate and missing exploration source material mean the experiment cannot test the claim. Conflict ratio should resolve after index rebuild. Status: hold_candidate, needs_redesign (pair replay with exploration mechanism). GOVERNANCE (2026-04-08): Manifests corrected, hold_candidate_resolve_conflict applied. Note: user confirmed V4 mechanisms being pulled into V3 scope -- MECH-165 substrate (reverse replay scheduler, exploration source material) should be planned for V3. - id: SD-017 title: "REE-v3 requires a minimal sleep-phase infrastructure: an SWS-analog phase (bidirectional hippocampus-to-cortex replay for schema/slot consolidation) and a REM-analog phase (causal attribution replay for slot-filling from co-correlational evidence ordering). Without these offline phases, hippocampal attribution mapping cannot converge and context representations remain globally undifferentiated." claim_type: design_decision subject: sleep_phase.minimal_sleep_infrastructure_v3 polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-092 - ARC-038 - ARC-007 - SD-014 location: docs/architecture/sd_017_sleep_phase_architecture.md notes: > Registered 2026-04-05. Significant roadmap change: previously all sleep phases were V4 scope (offline_phases.md, v3_v4_transition_boundary.md). Key insight: hippocampal attribution mapping -- the core E3 function -- requires offline consolidation phases to converge. Without them, context representations remain undifferentiated (cosine_sim -> 1.0; Law 2016 shows hippocampal context stabilisation requires ~3 interleaved sessions). Slot structure must exist before slot-filling yields reliable attribution signal (Sanders 2020 Bayesian hidden-state inference). The monostrategy failure mode (agent navigates like a fish swimming the same route) is not a V4 problem -- it appears in V3 and blocks E3's core attribution mapping function. V3 minimal implementation: (a) SWS-analog -- periodic hippocampus-to-cortex replay pass updating schema/context templates (distinct from MECH-092 which is hippocampus-internal waking quiescence replay); (b) REM-analog -- causal attribution replay filling context slots with co-correlational estimates from recent trajectory evidence, using information-direction switching. This is architecturally distinct from the full V4 sleep machinery: MECH-120 (SHY synaptic homeostasis), MECH-121 (NREM SWR+spindles), MECH-122 (thalamo-cortical spindles), MECH-123 (REM precision recalibration). Those remain V4. SD-017 is the V3 scaffold that enables E3 goal and harm streams to produce useful attribution maps, without which ree-v3 cannot consolidate contextual knowledge and remains behaviourally impoverished regardless of the quality of the online waking substrate. - id: ARC-045 title: "Bidirectional hippocampus-neocortex information flow is architecturally necessary for attribution mapping convergence: the SWS-direction (hippocampus to cortex) installs context schema structure; waking and REM-direction reactivations fill those structures with causal attribution estimates. Without bidirectional flow, context representations remain globally undifferentiated despite locally coherent online encoding." claim_type: architectural_commitment subject: hippocampus.bidirectional_information_flow polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-017 - MECH-092 - ARC-038 - MECH-166 notes: > Registered 2026-04-05. Biological grounding: hippocampal-to-neocortical information flow during slow-wave sleep (Diekelmann & Born 2010) is the canonical schema-installation pathway. Waking and REM reactivation then reads from and elaborates installed schemas. The bidirectional nature is not incidental -- it reflects a functional division: the hippocampus proposes structure; the neocortex receives, stabilises, and provides context feedback. ARC-045 claims bidirectionality is necessary (not merely useful) for attribution mapping convergence, because unidirectional online encoding alone cannot produce the stable context representations required for reliable causal attribution. Without the hippocampus-to-cortex pass installing schema attractors, the filling pass has no stable slots to fill. Experimental implication: an agent with bidirectional offline flow should show cosine_sim < 0.95 (differentiated contexts) after sleep phases; one with only waking online encoding remains at cosine_sim -> 1.0 regardless of training duration. EXQ-239 (MECH-153) approximates this via the supervised signal proxy -- the lambda=0.5 supervised pass is a weak stand-in for the SWS-direction schema-installation pass that SD-017 will properly implement. Bayesian framing (INV-044): the hip->cx (SWS) direction is the generative model installation pass -- hippocampus proposes structure, neocortex stabilises as attractors. The cx->hip (REM) direction is the recognition pass -- prediction error signals update the posterior. This bidirectional duality is structurally isomorphic to the generative/recognition architecture of hierarchical Bayesian models (predictive coding, Friston 2005-2010), implemented via information direction switching rather than explicit probability distributions. - id: MECH-166 title: "Context slot-formation (schema abstraction: which distinctions between contexts are worth representing) and slot-filling (causal attribution: which evidence populates each slot) are computationally separated functions that cannot be co-computed in a single online pass. Slot structure must be consolidated during an SWS-analog phase before slot-filling during a REM-analog phase can yield reliable attribution signal." claim_type: mechanism_hypothesis subject: hippocampus.slot_formation_filling_temporal_separation polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-017 - ARC-045 - MECH-092 - INV-044 notes: > Registered 2026-04-05. Grounded in Sanders, Wilson & Gershman 2020 (eLife): context discrimination at inference time is bounded by the quality of the learned hidden-state prior. The prior must be learned (slot-formation) before posterior inference (slot-filling) can be informative. In REE terms: the context templates (which terrain/situation types exist) must be consolidated as stable attractors before the agent can reliably attribute observed evidence to the correct context slot. Relationship to MECH-165: MECH-165 addresses replay content diversity (forward/reverse balance) during slot-filling. MECH-166 is the prerequisite: the slots must exist and be differentiated before diversity of replay content matters. A system with MECH-165 but not MECH-166 would have diverse replay content but no stable context structure to organise it into. The three waking/sleep information phases map as: (1) waking -- gather what is out there (E1 world model, online experience); (2) SWS-analog -- form the slots (schema abstraction, which distinctions matter); (3) REM-analog -- fill the slots (causal attribution from co-correlations and evidence ordering). Experimental implication: EXQ-239 (MECH-153) provides an indirect test. A direct test requires implementing the SWS-analog pass and comparing attribution map quality (context cosine_sim, terrain classification accuracy) with vs without it. - id: MECH-167 title: "z_harm_a (harm-driven affective accumulation) and drive_level (energy-depletion homeostatic deviation) instantiate the same computational motif -- slow interoceptive integration converging on ACC-type processing -- without sharing information or overlapping downstream gates. The motif predicts: (a) similar temporal dynamics (slow EMA-type accumulation), (b) orthogonal signal content (cos_sim -> 0), (c) distinct downstream gates: z_harm_a gates harm evaluation; drive_level gates z_goal seeding." claim_type: mechanism_hypothesis subject: homeostatic_accumulation.computational_motif polarity: asserts status: candidate implementation_phase: v3 depends_on: - SD-011 - SD-012 notes: > Registered 2026-04-05. Biological grounding: Craig 2003 (Trends Neurosci) groups pain, hunger, thirst, and temperature under a common lamina-I homeostatic pathway converging on ACC via VMpo. This creates an apparent scope collision between z_harm_a and drive_level. MECH-167 resolves this: the shared substrate reflects a repeated computational motif (slow interoceptive integration) not information sharing. The analogy: canonical microcircuit motifs appear across cortical areas without those areas sharing content. The two accumulators are separated by: (a) input source (harm exposure vs. energy depletion), (b) temporal dynamics (harm accumulates on contact; energy depletes continuously), (c) downstream gate (harm evaluation vs. z_goal seeding). Experimental prediction: z_harm_a and drive_level should show similar autocorrelation curves (slow decay) but near-zero cross-correlation (orthogonal content). A direct test regresses one against the other across a long episode and expects R^2 < 0.05. Generalisation: any future homeostatic signal (novelty depletion, social contact deficit) should follow the same integration architecture -- slow EMA, ACC-type convergence, distinct downstream gate -- instantiating the same motif without information sharing. This is a framework prediction, not just a description of the existing two accumulators. Empirical Bayes framing (INV-044): the slow EMA accumulators instantiate empirical Bayes over need states. The accumulated baseline is the running prior expectation -- P(harm_level) or P(drive_level). Deviations from baseline are precision-weighted signals: a harm event surprising relative to the prior carries more weight than one that confirms expectation. INV-044's prior-before-posterior requirement applies: a newly initialised accumulator with no history is a flat prior -- deviations are maximally uninformative until the prior has converged. - id: INV-044 title: "Approximate Bayesian contextual inference is architecturally impossible to co-compute with online encoding: prior construction (schema formation -- which context distinctions are structurally relevant) must precede posterior inference (attribution -- which evidence belongs to which context slot); a system attempting both online produces a degenerate prior that makes attribution uninformative regardless of training duration." claim_type: invariant subject: cognition.bayesian_prior_before_posterior_necessity polarity: asserts status: candidate notes: > Registered 2026-04-05. Grounded in Sanders, Wilson & Gershman 2020 (eLife doi:10.7554/eLife.51140): context discrimination accuracy at inference time is mathematically bounded by the quality of the learned hidden-state prior. The prior must be learned before posterior inference can be informative. This is not a heuristic -- a flat or degenerate prior makes the posterior unresponsive to evidence; no amount of additional evidence can compensate for a missing prior structure. In REE terms: slot-formation (SWS-analog pass) builds the prior -- the schema of which context distinctions are structurally relevant. Slot-filling (REM-analog pass) performs the posterior -- which evidence belongs to which context slot. An agent attempting both online accumulates evidence against a flat prior and produces globally undifferentiated representations (cosine_sim -> 1.0), exactly as Law et al. 2016 and the MECH-153 failure mode document. Predictive coding connection (ARC-045): the SWS-direction (hip->cx) propagation is the generative model installation pass; the REM-direction carries the prediction error that updates the posterior. This is structurally the architecture of hierarchical Bayesian models (Friston 2005-2010) implemented through information direction switching, not explicit probability distributions. Empirical Bayes extension (MECH-167): homeostatic accumulators (z_harm_a, drive_level) instantiate the same principle at the need-state level. The slow EMA baseline is the prior; deviations are likelihood-weighted signals. The accumulator must have converged before deviation signals are interpretable. This invariant is what makes SD-017 a V3 necessity rather than a V4 optimisation: not "sleep phases would help" but "sleep phases are required for approximate Bayesian contextual inference to be architecturally coherent." Mechanistic instantiation: MECH-166 (slot-formation/filling temporal separation), SD-017 (SWS-analog + REM-analog design decision), ARC-045 (bidirectional flow). - id: INV-045 title: "The biological NREM-then-REM sleep phase sequence is the correct computational order, derivable from the logical dependencies between the failure modes each phase addresses; this ordering is not a contingent biological fact but an engineering necessity, and it serves directly as the V4 rewiring specification: each phase specifies what structural change the REE substrate requires to address its failure mode." claim_type: invariant subject: sleep.phase_ordering_computational_necessity polarity: asserts status: candidate depends_on: - INV-044 - SD-017 - MECH-120 - MECH-121 - MECH-122 - MECH-123 notes: > Registered 2026-04-05. The V4 sub-phase sequence is: (0) sensory gating, (1) SHY normalisation, (2) NREM schema replay, (3) spindle coordination, (4) REM precision recalibration. Each phase is in this position because of what must be true before it can run: Phase 0 must be first: new input corrupts in-progress consolidation because schemas shift during installation. Nothing downstream can run while perception is live. Rewiring: input gate on E1 latent stack; ThetaBuffer paused for new observations. Phase 1 (MECH-120) must precede replay: replaying into a weight landscape still dominated by recent high-salience experiences would reinforce the dominant trace rather than consolidate diverse content. Homeostasis must flatten attractors before replay repopulates them. Rewiring: normalisation pass over E1/E2/E3 weights decaying toward mean; dominant attractor suppression. Phase 2 (MECH-121 + MECH-165) must precede REM: INV-044 applies -- you cannot fill context slots that do not yet exist. Schema installation must produce stable context attractors before causal attribution replay can assign evidence to them. Replay diversity (MECH-165) must be in this phase because diverse path content must enter the schema, not just the dominant trajectory. Rewiring: hippocampus-to-cortex directed replay with balanced scheduler (forward + reverse + non-dominant paths); ContextMemory templates updated. Phase 3 (MECH-122) packages E1 state for bidirectional consolidation. ThetaBuffer has been waking-only in V3; it must become bidirectional to allow E1 updates to be consolidated into E3 in the reverse direction. Rewiring: ThetaBuffer gains a reverse-direction mode; theta-packaged E1 updates transferred to hippocampus for long-horizon integration. Phase 4 (MECH-123) must be last: resetting precision priors before replay (phases 2-3) would change how that evidence is weighted during replay -- the prior recalibration must follow, not precede, the content consolidation passes. Rewiring: z_beta suppressed (aminergic suppression analog); E1 runs unconstrained (no harm gate, unconstrained simulation); commitment_threshold and precision_ema_alpha recalibrated from the episode's natural range. V4 engineering implication: this invariant means the V4 sleep phase controller is not arbitrary in its sequencing. A controller that ran phases in a different order would either fail to consolidate (wrong order of schema vs attribution) or corrupt the weight landscape (replay before homeostasis). The ordering is a constraint on the implementation, not a design choice. See offline_phases.md for the rewiring table. # ---- Claims registered 2026-04-05: dementia as attribution pipeline failure ---- - id: INV-046 title: "Dementia is computationally a progressive failure of contextual attribution rather than primarily a memory storage failure: episodic memory loss is downstream of progressive collapse of the offline consolidation pipeline that maintains the context attribution scaffold." claim_type: invariant subject: dementia.attribution_pipeline_failure_reframe polarity: asserts status: candidate depends_on: - INV-045 - INV-044 - MECH-166 - SD-017 notes: > Registered 2026-04-05. The conventional framing of dementia as memory storage failure mislocates the primary lesion. The offline consolidation pipeline (SWS-analog schema installation, REM-analog slot-filling, MECH-120 homeostasis, MECH-123 precision recalibration) is the infrastructure that makes contextual attribution possible. Progressive collapse of this pipeline produces progressive failure of contextual attribution -- which manifests as the clinical syndrome of dementia. Memory loss is not the primary failure; it is the observable consequence of the agent losing the ability to attribute the meaning of its experiences to the correct context slots. Relationship to existing claims: MECH-124 describes consolidation-mediated option-space contraction (Walker PTSD analog -- imbalanced replay produces over-consolidation of harm traces). INV-046 describes a distinct failure mode: insufficient consolidation cycles producing progressive dissolution of the context scaffold itself. PTSD over-consolidates; dementia under- consolidates. The failure mode, direction, and symptom profile are different. Derivation: follows directly from the REE attribution pipeline. An agent whose offline phases degrade progressively will first lose precision in contextual attributions (Phase 4 failure), then lose new episodic slot- filling (Phase 2-3 failure), then lose the context slots themselves (Phase 1-2 failure), then lose global contextual orientation (Phase 0 equivalent). This is not an analogy to dementia -- it is what the computational account predicts would happen to any agent with progressively degrading offline consolidation cycles. - id: INV-047 title: "The specific staged cognitive decline profile of Alzheimer's disease -- early precision/confidence loss, then episodic memory failure, then contextual identity loss, then global disorientation -- follows the reverse of the sleep phase dependency ordering and is not explained by diffuse amyloid damage alone, which predicts uniform degradation rather than a staged profile." claim_type: invariant subject: dementia.staged_decline_phase_dependency_prediction polarity: asserts status: candidate depends_on: - INV-046 - INV-045 - MECH-168 notes: > Registered 2026-04-05. The glymphatic clearance hypothesis (amyloid accumulation without sleep; Xie/Nedergaard 2013) predicts substrate damage that should be diffuse across the consolidation pipeline, not staged. But the clinical presentation of Alzheimer's is highly staged: new episodic learning fails before remote memory; contextual identity (who is this person, what does this place mean) fails before procedural skill; global orientation fails last. Diffuse damage cannot explain this profile. The attribution pipeline dependency ordering predicts exactly this staging: Phase 4 (precision recalibration, MECH-123) is the most downstream and most vulnerable -- its failure is the earliest detectable symptom (the "something feels off" prodrome of MCI, reduced confidence in attributions before overt memory failure). Phase 2-3 (slot-filling, MECH-121) failure produces the classic amnesic profile: intact remote schemas, absent new consolidation. Phase 1-2 (slot-formation, MECH-120) failure destabilises existing schemas -- middle-stage loss of contextual identity. Phase 0 equivalent failure produces global contextual collapse. This is a specific, falsifiable prediction: the temporal ordering of symptom emergence should map onto the reverse of the phase dependency graph, not onto the spatial distribution of amyloid/tau pathology. Neuroanatomical note: hippocampal neurodegeneration (early in Alzheimer's) selectively impairs the hip->cx information direction (SWS-analog, schema installation) before disrupting the cx->hip direction (waking encoding), predicting that schema formation fails before schema retrieval -- consistent with the amnesic profile of early Alzheimer's (new learning fails before remote memory access). - id: MECH-168 title: "Progressive sleep deprivation or neurodegeneration causes staged failure of the offline consolidation pipeline in reverse dependency order: precision recalibration (MECH-123/Phase 4) degrades first, producing early MCI features; slot-filling (MECH-121/Phase 2-3) next, producing the classic early Alzheimer's episodic profile (intact remote, absent new); slot-formation (MECH-120/Phase 1-2) last, producing middle-stage contextual identity loss and schema destabilisation; global context collapse last." claim_type: mechanism_hypothesis subject: dementia.consolidation_pipeline_staged_failure polarity: asserts status: candidate depends_on: - INV-046 - INV-047 - INV-045 - MECH-120 - MECH-121 - MECH-123 - SD-017 notes: > Registered 2026-04-05. The mechanism: each phase of the offline consolidation pipeline depends on the phases upstream of it. When the pipeline degrades progressively (via sleep deprivation, neurodegeneration, or disrupted sleep architecture), the most downstream phases fail first because they are furthest from the biological substrate and depend on the most intact upstream machinery. Phase 4 (MECH-123 precision recalibration) fails first because it requires the full pipeline to have run -- any degradation in earlier phases propagates to imprecise precision calibration. Clinical correlate: early MCI, reduced confidence in attributions, increased uncertainty about whether current experience matches prior context, the subjective sense of unreliability before overt memory failure. Phase 2-3 (MECH-121 slot-filling) fails next. New episodic experiences stop consolidating into existing context schemas. Recent events, new faces, and new environments fail to attach to the correct context slots. Remote memories (consolidated before pipeline degradation) remain accessible. Clinical correlate: classic early Alzheimer's amnesic profile. Phase 1-2 (MECH-120 + MECH-121 slot-formation) fails last. The existing schema structure begins to destabilise -- not just failure to add new content, but erosion of the context attractors themselves. Who is this person? What does this place mean? The agent loses the slots, not just their contents. Clinical correlate: middle-stage Alzheimer's, loss of contextual identity. Global collapse: with no stable context scaffold remaining, attribution is impossible. The agent is oriented to nothing. Clinical correlate: late-stage dementia. Distinction from MECH-124 (PTSD failure mode): MECH-124 is consolidation-mediated option-space contraction via imbalanced replay -- too much consolidation of harm traces, too little of alternative paths. MECH-168 is insufficient consolidation cycles producing progressive dissolution of the context scaffold. PTSD consolidates the wrong content; dementia fails to consolidate at all. - id: MECH-169 title: "The glymphatic clearance hypothesis (amyloid-beta accumulation via insufficient sleep-phase clearance) and the attribution pipeline account (consolidation failure via disrupted offline phases) are mechanistically complementary accounts of sleep-deprivation-induced dementia: the former explains why structural substrate damage accumulates; the latter explains why cognitive symptoms are staged and context-specific rather than diffuse and uniform." claim_type: mechanism_hypothesis subject: dementia.glymphatic_attribution_complementarity polarity: asserts status: candidate depends_on: - INV-046 - INV-047 - MECH-168 notes: > Registered 2026-04-05. The glymphatic hypothesis (Xie et al. 2013, Science: glymphatic system clears metabolic waste including amyloid-beta predominantly during sleep; sleep deprivation leads to amyloid accumulation) is well-supported and explains the substrate damage pathway. MECH-169 does not challenge it -- it extends it. The glymphatic account leaves two things unexplained: (1) why the cognitive symptom profile is staged rather than uniform given that amyloid accumulates diffusely; (2) why the specific profile is contextual-attribution failure rather than, say, motor or sensory failure. MECH-168/INV-047 provide the explanation: the attribution pipeline has a dependency ordering that produces staged failure even under diffuse substrate damage, because downstream phases fail before upstream phases. The two accounts are therefore complementary: - Glymphatic: explains structural damage accumulation (why sleep deprivation causes Alzheimer's risk at the substrate level) - Attribution pipeline: explains symptom profile shape (why the cognitive decline is staged, contextual, and follows the reverse of the phase dependency ordering) A complete mechanistic account of sleep deprivation-induced dementia requires both. The glymphatic account without the attribution pipeline account cannot predict the clinical phenotype. The attribution pipeline account without the glymphatic account cannot explain why sleep deprivation specifically (rather than other forms of neural disruption) is the primary risk factor. - id: MECH-170 title: "Sleep architecture restoration in early MCI should produce dissociated cognitive improvement: slot-filling (episodic attribution, MECH-121 analog) should improve before slot-formation (contextual semantic, MECH-120 analog), because the pipeline rebuilds in dependency order. This distinguishes the attribution account from accounts predicting uniform memory improvement with sleep restoration." claim_type: mechanism_hypothesis subject: dementia.sleep_restoration_dissociated_recovery_prediction polarity: asserts status: candidate depends_on: - MECH-168 - INV-047 - MECH-120 - MECH-121 notes: > Registered 2026-04-05. Testable prediction: if the attribution pipeline account is correct, sleep architecture restoration in early MCI patients should not produce uniform improvement across memory domains. The pipeline rebuilds in dependency order: Phase 4 (precision) restores first, then Phase 2-3 (slot-filling/episodic), then Phase 1-2 (slot-formation/semantic). Therefore the predicted sequence of improvement with sleep intervention is: (1) improved confidence and precision in contextual attributions (early, subtle, may require sensitive testing); (2) improved new episodic memory consolidation (formation of new slot-filling associations); (3) improved contextual semantic memory (only if intervention is sustained long enough for schema rebuilding). This dissociation distinguishes the attribution account from: - General memory consolidation accounts, which predict proportional improvement across all memory types - Glymphatic accounts, which predict improvement proportional to reduced amyloid load (uniform across domains) Clinical operationalisation: prospective sleep intervention study in early MCI, measuring episodic (new word pairs, face-name) vs semantic (contextual categorisation, situation identity) memory at intervals. Attribution account predicts episodic recovery precedes semantic recovery. The prediction is specific enough to be falsifiable within a single randomised controlled trial. - id: MECH-171 title: "Progressive neurodegeneration creates a vicious cycle through sleep disruption: early attribution pipeline failure disrupts sleep architecture, which accelerates consolidation pipeline failure, which accelerates attribution failure. This predicts disproportionate disease-modifying effect of early sleep intervention, because it breaks the cycle before the schema scaffold collapses." claim_type: mechanism_hypothesis subject: dementia.sleep_disruption_vicious_cycle polarity: asserts status: candidate depends_on: - MECH-168 - INV-046 notes: > Registered 2026-04-05. The vicious cycle operates in both directions: Forward: sleep deprivation -> consolidation failure -> attribution failure -> behavioural disorientation -> increased arousal/anxiety -> further sleep disruption. Reverse: early neurodegeneration (hippocampal volume loss) disrupts the hip->cx information pathway (SWS-analog direction) -> schema installation fails -> REM-analog slot-filling operates against an unstable prior -> precision recalibration is less effective -> waking behaviour is less well-calibrated -> increased cognitive effort and arousal -> further sleep disruption. The vicious cycle predicts: (1) sleep disruption should be both a risk factor for and an early symptom of Alzheimer's (consistent with Lucey et al. 2017, JAMA Neurology: sleep disruption precedes amyloid deposition by years); (2) the cycle amplifies -- early mild disruption becomes severe disruption as the attribution scaffold degrades; (3) early intervention (before schema scaffold collapse) has disproportionately large effect relative to late intervention, because breaking the cycle early prevents amplification. Clinical implication: this suggests a critical window for sleep intervention in MCI/pre-dementia that is much earlier than current therapeutic targeting. Once the schema scaffold has destabilised (middle-stage), sleep restoration alone cannot rebuild it -- the slots that would be filled by recovered REM-analog function no longer exist. - id: MECH-172 title: "Preservation of habit and procedural memory in early Alzheimer's disease, while episodic and contextual memory fail, is architecturally predicted by the separation of the model-free habit system (MECH-163 D1/D2, SNc) from the contextual attribution system (E3 hippocampal, requiring offline consolidation phases): the habit system does not require offline consolidation to function, so its relative preservation under early consolidation pipeline failure is not incidental but follows from the architectural separation." claim_type: mechanism_hypothesis subject: dementia.habit_system_sparing_architectural_prediction polarity: asserts status: candidate depends_on: - INV-046 - MECH-168 - MECH-163 - SD-017 notes: > Registered 2026-04-05. Clinical observation: patients with early to middle-stage Alzheimer's disease typically retain procedural and habit memory (riding a bicycle, playing a musical instrument, daily routines) long after episodic and contextual memory has substantially failed. This is conventionally attributed to anatomical sparing of the basal ganglia and striatum relative to hippocampal/entorhinal neurodegeneration. The attribution account provides a deeper computational explanation: the two systems have different consolidation requirements. The model-free habit system (MECH-163 D1/D2, analogous to SNc-striatal loops) learns via online reinforcement and does not require offline consolidation phases to maintain its representations -- the weights are updated during waking and do not depend on hippocampus-to-cortex schema installation. The contextual attribution system (E3 hippocampal, residue field) requires offline consolidation to maintain differentiated context representations. When the consolidation pipeline degrades, the habit system continues to function while the attribution system fails -- producing exactly the selective sparing observed clinically. This is a double dissociation that confirms the architectural separation: (1) attribution system fails while habit system is spared (early AD); (2) in pure habit system disorders (Huntington's, striatal lesions), procedural memory fails while episodic memory is relatively spared. The REE account predicts both dissociations from first principles. # ---- Claims registered 2026-04-05: medication effects on sleep architecture as attribution pipeline disruption ---- - id: INV-048 title: "Any pharmacological agent that disrupts sleep architecture produces equivalent attribution pipeline degradation to behavioral sleep deprivation, proportional to which phase(s) are disrupted and the duration of exposure -- the pipeline is mechanism-agnostic and responds only to phase fidelity." claim_type: invariant subject: sleep.pharmacological_disruption_pipeline_equivalence polarity: asserts status: candidate depends_on: - INV-045 - INV-046 - MECH-120 - MECH-121 - MECH-123 notes: > Registered 2026-04-05. The attribution pipeline (INV-045/INV-046) makes no mechanistic distinction between sleep deprivation caused by behavioral factors and sleep deprivation caused by pharmacological suppression of sleep phases. The relevant variable is phase fidelity -- whether each consolidation phase runs with adequate depth, duration, and sequence integrity -- not the cause of its disruption. This equivalence principle grounds all downstream pharmacological predictions (MECH-173-177): medication review becomes a pipeline audit. For each prescribed medication, the relevant question is: which pipeline phases does this agent suppress, by how much, over how long? The grounding follows directly from INV-045 (phase ordering is computationally necessary) and INV-046 (dementia is attribution pipeline failure). If the pipeline requires phase fidelity, and medications can disrupt phase fidelity, then medications can cause pipeline degradation via the same mechanism as behavioral sleep loss. Epidemiological support: benzodiazepine and anticholinergic use in older adults are associated with increased dementia incidence (Billioti de Gage et al. 2014 BMJ; Coupland et al. 2019 JAMA Internal Medicine), consistent with pharmacological pipeline disruption over years. - id: MECH-173 title: "Medications that suppress REM sleep (anticholinergics, MAOIs, most antidepressants at therapeutic doses, benzodiazepines) selectively impair MECH-123 precision recalibration -- the most downstream pipeline phase -- and thereby accelerate the earliest dementia prodrome: overconfident contextual attributions before overt memory loss." claim_type: mechanism_hypothesis subject: pharmacology.rem_suppression_phase4_disruption polarity: asserts status: candidate depends_on: - INV-048 - MECH-123 - INV-047 - MECH-168 notes: > Registered 2026-04-05. REM sleep is the substrate of MECH-123 (precision recalibration): during REM, decision thresholds and prior confidence weights are reset, allowing the waking system to update its uncertainty estimates. Pharmacological REM suppression deprives this recalibration of its substrate. Because Phase 4 (precision recalibration) fails first in natural neurodegeneration (INV-047), pharmacological REM suppression selectively targets the most vulnerable phase, accelerating the earliest prodrome: the agent continues acting with high confidence while the accuracy of contextual attributions degrades -- the combination characteristic of early MCI (subjective normalcy, objective deficit on sensitive testing). Key REM-suppressing medications by mechanism: (1) Anticholinergics: ACh (from nucleus basalis of Meynert and brainstem PGO wave generators) is the primary neurochemical driver of REM; any muscarinic blockade suppresses REM directly. (2) MAOIs: profound REM suppression via noradrenergic/serotonergic excess; clinically, REM is almost abolished at therapeutic doses. (3) SSRIs/SNRIs: REM suppression via serotonergic excess acting on 5-HT2A REM-off neurons (paroxetine most REM-suppressive due to combined anticholinergic action; venlafaxine and duloxetine moderate). (4) TCAs: REM suppression via combined anticholinergic + serotonergic effects; amitriptyline and clomipramine most potent. (5) Benzodiazepines: reduce REM% via preferential Stage 2 enhancement at the expense of SWS and REM. Prediction: cumulative REM-suppressing medication years should predict MCI conversion, detectable as precision failure on laboratory tasks before clinical diagnosis. This is testable via retrospective pharmacoepidemiology with neuropsychological battery data. - id: MECH-174 title: "GABAergic sleep aids (benzodiazepines and Z-drugs: zolpidem, zopiclone, zaleplon) suppress SWS depth and continuity at therapeutic doses, impairing MECH-120 synaptic homeostasis and MECH-121 schema replay, producing context representations that remain globally undifferentiated -- the same failure mode as SD-017 without any SWS-analog." claim_type: mechanism_hypothesis subject: pharmacology.sws_suppression_phase12_disruption polarity: asserts status: candidate depends_on: - INV-048 - MECH-120 - MECH-121 - SD-017 notes: > Registered 2026-04-05. GABA-A positive allosteric modulators (classical benzodiazepines: diazepam, lorazepam, temazepam, nitrazepam; and Z-drugs: zolpidem, zopiclone, zaleplon) bind to benzodiazepine-sensitive GABA-A receptor subunits and produce sleep via global CNS inhibition. This global inhibition suppresses SWS in two distinct ways: (1) Direct SWS depth suppression: slow oscillation amplitude is reduced because the hypnotic state is produced pharmacologically rather than via homeostatic pressure -- the homeostatic drive that normally generates deep SWS is partially bypassed. This impairs MECH-120 (SHY synaptic homeostasis): the normalisation of dominant attractors requires adequate slow-wave power to drive the homeostatic reset. (2) SWS continuity disruption: GABAergic hypnotics produce fragmented slow-wave episodes rather than the sustained SWS blocks required for MECH-121 (schema replay via SWR sequences). Schema consolidation depends on sustained hippocampal SWR sequences coordinated with cortical slow oscillations (Staresina et al. 2015, Diekelmann & Born 2010); fragmented SWS produces fragmented and incomplete replay. Note on Z-drug distinction: Z-drugs preferentially bind alpha-1 subunit GABA-A receptors and enhance Stage 2 spindles (which may support MECH-122 spindle coordination for already-consolidated items). However, they still reduce SWS homeostatic depth because they produce sleep via direct pharmacological inhibition rather than via accumulated homeostatic pressure. The net SWS homeostasis impairment is therefore shared with classical benzodiazepines, even if spindle coordination is partially preserved. Prediction: long-term benzodiazepine/Z-drug users should show reduced SWS% on polysomnography and slower context-discrimination learning on tasks requiring schema formation, with episodic retrieval relatively more preserved than new context learning -- the signature of Phase 1/2 failure. - id: MECH-175 title: "Anticholinergic medications confer dementia risk via two independent and potentially additive pathways: (1) REM suppression (ACh is required for REM generation via nucleus basalis and brainstem PGO generators -- muscarinic blockade directly suppresses REM, impairing MECH-123); and (2) cholinergic deficit mimicry (AD is characterised by basal forebrain cholinergic loss; anticholinergic burden creates a pharmacological analog of this deficit during waking hours, independently degrading attention and encoding)." claim_type: mechanism_hypothesis subject: pharmacology.anticholinergic_dual_pathway_dementia_risk polarity: asserts status: candidate depends_on: - MECH-173 - INV-048 notes: > Registered 2026-04-05. The dual-pathway account distinguishes two independent mechanistic contributions of anticholinergic burden to dementia risk: Pathway 1 -- nocturnal REM suppression (pipeline disruption): ACh is the primary neurochemical driver of REM. Muscarinic M1/M2 receptors on PGO (ponto-geniculo-occipital) wave generating neurons and nucleus basalis of Meynert project to cortex and drive REM-state cortical desynchronisation. Anticholinergic blockade directly suppresses PGO waves, suppressing REM initiation and maintenance. This degrades MECH-123 precision recalibration nightly, compounding over years. Pathway 2 -- diurnal cholinergic deficit mimicry (encoding disruption): Alzheimer's disease is characterised by progressive loss of basal forebrain cholinergic neurons (nucleus basalis of Meynert). The cholinergic hypothesis of Alzheimer's (Bartus et al. 1982) established that this deficit is the primary driver of early attentional and encoding failure. Anticholinergic medications produce an iatrogenic cholinergic deficit during waking hours -- a pharmacological analog of the AD lesion itself. This independently impairs attention, encoding, and the cortical desynchronisation required for active learning. Additive prediction: anticholinergic burden scores (ACB scale, Anticholinergic Risk Scale, German Anticholinergic Burden score) should predict dementia conversion better than total medication count or total sleep medication count, because they capture both pathway contributions. The dose-response relationship should be approximately additive if the two pathways are mechanistically independent. This is testable via mediation analysis: if anticholinergic burden -> dementia risk is partially mediated via REM-suppression-years, the nocturnal pathway is confirmed as independent. High-burden medications by indication class: Bladder/urinary: oxybutynin (high), solifenacin (moderate), tolterodine (moderate), darifenacin (high). OTC antihistamines: diphenhydramine (high), chlorphenamine (moderate). TCAs: amitriptyline (high), clomipramine (high), doxepin (high), imipramine (high). Antipsychotics: chlorpromazine (high), olanzapine (moderate), clozapine (high), quetiapine (low-moderate). SSRIs: paroxetine (moderate -- outlier in SSRI class due to potent muscarinic blockade; other SSRIs are low burden). Epidemiological support: Coupland et al. 2019 (JAMA Internal Medicine), Gray et al. 2015 (JAMA Internal Medicine), Khachaturian et al. 2006 (Archives of Neurology) -- consistent dose-response anticholinergic burden-dementia association replicated across multiple cohorts and healthcare systems. - id: MECH-176 title: "Orexin receptor antagonists (suvorexant, lemborexant) are predicted to be architecture-preserving sleep aids: they suppress wake-promoting orexin drive without directly compressing sleep phase machinery, allowing the brain's own homeostatic sleep architecture to unfold, and should therefore preserve or improve SWS/REM ratio relative to GABAergic alternatives at equi-efficacious doses." claim_type: mechanism_hypothesis subject: pharmacology.orexin_antagonist_architecture_preserving polarity: asserts status: candidate depends_on: - MECH-174 - INV-048 notes: > Registered 2026-04-05. GABAergic sleep aids produce sleep by actively suppressing the nervous system globally. This bypasses homeostatic pressure, reducing SWS depth and compressing REM by competing with the endogenous machinery (MECH-174). Orexin receptor antagonists (DORAs: dual orexin receptor antagonists -- suvorexant, lemborexant) work by a different mechanism: they block the wake-promoting orexin/hypocretin signal from lateral hypothalamus, permissively allowing the homeostatic sleep pressure to express itself without pharmacological interference with the sleep machinery. The key distinction: GABAergic agents add an active suppressive signal; orexin antagonists remove an active wake-promoting signal. The result is that homeostatic SWS generation is not competed with, and REM cycles emerge from endogenous REM-permitting neurochemistry (reduced monoamines, cholinergic rebound) rather than from pharmacological REM-off suppression. PSG evidence: multiple head-to-head and placebo comparisons show orexin antagonists maintain or slightly increase REM% and do not suppress SWS% below placebo levels (Herring et al. 2016, SLEEP; Kishi et al. 2020 meta-analysis). This contrasts with zolpidem (SWS suppression) and temazepam (REM + SWS compression). Testable prediction: in older adults at risk for MCI, orexin antagonist users should show (1) preserved SWS% and REM% on polysomnography, (2) slower episodic memory decline compared to benzodiazepine/Z-drug users at matched total sleep time, and (3) better new context discrimination performance (SD-017 analog in humans: ability to distinguish similar but distinct contexts in list learning tasks). Clinical implication: when a sleep aid is indicated in an older adult at MCI risk, orexin antagonists should be preferred over GABAergic alternatives specifically on dementia risk grounds, independent of other safety considerations (falls, dependence, tolerance). - id: MECH-177 title: "Any pharmacological intervention that durably restores the NREM-to-REM sequence and preserves phase durations in MCI or early Alzheimer's disease is predicted to be disease-modifying via pipeline protection -- upstream of irreversible neurodegeneration, breaking the MECH-171 vicious cycle before schema scaffold collapse, with predicted effect size larger in early MCI than established AD." claim_type: mechanism_hypothesis subject: pharmacology.sleep_architecture_restoration_disease_modifying polarity: asserts status: candidate depends_on: - MECH-176 - MECH-171 - INV-048 - MECH-170 notes: > Registered 2026-04-05. Pipeline protection mechanism: the MECH-171 vicious cycle (sleep disruption -> consolidation failure -> attribution failure -> increased arousal -> further sleep disruption) is the amplification mechanism for dementia progression. Breaking the cycle pharmacologically -- by restoring phase-adequate sleep -- interrupts amplification before the schema scaffold collapses. This is a disease-modifying effect, not symptomatic treatment, because it preserves the infrastructure that contextual attribution depends on. Critical window: MECH-171 predicts that early intervention (MCI stage, before schema scaffold destabilises) produces disproportionately large effects relative to late intervention (established AD, after context slots no longer exist). This is consistent with the general principle that disease-modifying interventions in AD are stage-dependent. Candidates by predicted mechanism and evidence tier: (1) Suvorexant / lemborexant (MECH-176): architecture-preserving via orexin antagonism; preserves SWS + REM. Active clinical investigation (Herring et al. 2020 Alzheimer's Research and Therapy). (2) Ramelteon / melatonin-extended-release: circadian restoration (Phase 0 in INV-045 terminology) improves architecture downstream. Melatonin-ER licensed in EU specifically for over-65 on architecture preservation grounds. Ramelteon: no direct REM/SWS compression. (3) Trazodone (low dose, 25-100mg): 5-HT2A antagonism increases SWS depth (Phase 1/2 enhancement); used in AD sleep disruption trials. Active RCT (McCleery & Sharpley 2020 Cochrane review). SWS-enhancing mechanism is non-GABAergic and non-anticholinergic -- neutral on REM. Provides Phase 1/2 benefit with minimal Phase 4 penalty. (4) Gabapentin / pregabalin: increase SWS depth via voltage-gated calcium channel modulation. Phase 1/2 enhancement. Caveat: tolerance develops for SWS effect; fall risk and cognitive sedation in older adults warrant caution. Not a first-line DPM candidate. (5) Low-dose doxepin (Silenor, 3-6mg): histamine H1 antagonism only at this dose; minimal muscarinic activity below 10mg. Enhances sleep continuity with spindle preservation (MECH-122 compatible). Not REM-suppressive at low dose. Contrast with amyloid-targeting drugs (lecanemab, donanemab): these target substrate damage accumulation (glymphatic pathway, MECH-169) rather than the symptom-progression pipeline. MECH-177 and MECH-169 predict additive benefit from combined pipeline protection + amyloid clearance, because they operate on mechanistically independent targets. A phase II trial comparing orexin antagonist + sleep architecture monitoring vs standard care in MCI would directly test MECH-177. Primary endpoint: rate of episodic memory decline (MECH-170 prediction: episodic decline slows before semantic/contextual effects are detectable). - id: Q-031 title: "Does the anticholinergic burden-dementia risk relationship track cumulative REM suppression duration (total ACh-blocked sleep years), total anticholinergic dose as a proxy for diurnal cholinergic deficit, or both as independent additive predictors -- and can mediation analysis in prospective cohorts distinguish the nocturnal (MECH-173) from the diurnal (MECH-175 pathway 2) contribution?" claim_type: open_question subject: pharmacology.anticholinergic_risk_rem_vs_dose_independent_predictors polarity: open status: open depends_on: - MECH-175 - MECH-173 notes: > Registered 2026-04-05. Distinguishing the two anticholinergic pathways has direct clinical implications: if the nocturnal REM suppression mechanism (Pathway 1) dominates, then switching to non-anticholinergic alternatives with the same indication (e.g., bladder: oxybutynin -> mirabegron, a selective beta-3 adrenergic agonist with no CNS anticholinergic burden) should substantially reduce dementia risk even without improving daytime cognition. If the diurnal cholinergic deficit mechanism (Pathway 2) dominates, mirabegron switch reduces risk via the daytime pathway, and sleep architecture monitoring becomes the key biomarker. If both are additive, both interventions are required. Methodology to test: prospective cohort with (1) annual PSG subset to quantify REM-suppression-years, (2) anticholinergic burden scoring at each visit, (3) dementia conversion endpoint. Structural equation model with REM-suppression-years and cumulative anticholinergic dose as parallel mediators of the anticholinergic burden -> dementia path. Current evidence is insufficient to distinguish the pathways because existing anticholinergic-dementia studies (Coupland 2019, Gray 2015) do not include PSG data, making it impossible to separate nocturnal from diurnal mechanisms. - id: Q-032 title: "Among sleep-promoting medications prescribed in older adults, does relative SWS/REM ratio preservation (measured by polysomnography) predict differential dementia outcomes independent of total sleep time -- and if so, does this validate phase-specific pipeline damage as the causal intermediate and provide a pharmacodynamic biomarker for disease-modifying trials?" claim_type: open_question subject: pharmacology.sws_rem_ratio_dementia_outcome_predictor polarity: open status: open depends_on: - INV-048 - MECH-173 - MECH-174 notes: > Registered 2026-04-05. Total sleep time is the conventional endpoint for sleep intervention trials. If pipeline phase fidelity is the causal intermediate (INV-048), then total sleep time is an insufficient surrogate -- two agents producing equal total sleep time with different SWS/REM profiles should produce different dementia outcomes. If confirmed, SWS% and REM% on PSG would become phase-specific pharmacodynamic biomarkers for disease-modifying trials, enabling mechanism-targeted trial design. Specifically: (1) SWS% as surrogate for Phases 1/2 fidelity (MECH-120/121) (2) REM% as surrogate for Phase 4 fidelity (MECH-123) (3) Spindle density as surrogate for Phase 3 fidelity (MECH-122) Existing data: no direct comparison has used PSG-derived phase metrics as primary endpoints in a dementia conversion trial. The COSMOS-MIND trial (Mewton et al. 2023) and FINGER-type multimodal trials include sleep as a secondary domain but without PSG. A PSG-stratified pharmacoepidemiology study in existing EHR cohorts with linked PSG records (e.g., SHHS, MrOS Sleep Study) could test this retrospectively. - id: IMPL-026 title: "Medication classification by predicted attribution pipeline effect: a reference table mapping common sleep-affecting medication classes to their predicted phase-specific pipeline disruption, net dementia risk direction, and disease-modifying potential." claim_type: implementation_note subject: pharmacology.sleep_dementia_risk_classification polarity: records status: candidate depends_on: - MECH-173 - MECH-174 - MECH-175 - MECH-176 - MECH-177 - INV-048 notes: > Registered 2026-04-05. Reference classification for clinical use and trial design. Pipeline phase notation follows INV-045: Phase 0 = sensory gating (pre-sleep); Phase 1 = SHY homeostasis (MECH-120); Phase 2 = schema replay (MECH-121); Phase 3 = spindle coordination (MECH-122); Phase 4 = REM precision recalibration (MECH-123). Risk direction: WORSENS = net dementia risk increase; NEUTRAL = no significant pipeline disruption at therapeutic doses; IMPROVES = net pipeline protection or enhancement; CONDITIONAL = depends on dose, duration, or patient subtype. DPM = disease-progression-modifying potential. HIGH-RISK MEDICATIONS (significant pipeline disruption): Anticholinergics (all classes): Pipeline phases disrupted: Phase 4 (REM suppression) + diurnal cholinergic deficit (MECH-175). Risk: WORSENS. DPM potential: no (harmful). Priority: switch to non-anticholinergic alternatives wherever possible (mirabegron for bladder; SSRI/loratadine for antihistamine; other antidepressants for amitriptyline). Benzodiazepines: Pipeline phases disrupted: Phases 1/2 (SWS suppression) + Phase 4 (REM suppression) (MECH-173 + MECH-174). Risk: WORSENS. DPM potential: no. Priority: deprescribe; taper and replace with non-GABAergic alternatives. Z-drugs (zolpidem, zopiclone, zaleplon): Pipeline phases disrupted: Phase 1 (SWS homeostatic depth reduced) + Phase 4 (REM mildly reduced). Spindle enhancement may partially preserve Phase 3. Risk: WORSENS (less severe than classical BZDs but still harmful over years). DPM potential: no. Priority: replace with architecture-preserving alternatives. MAOIs: Pipeline phases disrupted: Phase 4 profound (nearly complete REM abolition at therapeutic doses). Risk: WORSENS. DPM potential: no. Use only when no alternatives exist; monitor with PSG if chronic use. SSRIs/SNRIs (most): Pipeline phases disrupted: Phase 4 moderate (REM suppression, especially paroxetine > venlafaxine > escitalopram/sertraline). Risk: WORSENS (moderate, chronic use). Exception: if treating depression/anxiety that is itself causing sleep disruption, net effect may be neutral or protective short-term. DPM potential: conditional (treating comorbid depression in MCI may have indirect benefit via MECH-171 cycle breaking). Prefer low-REM-burden SSRIs (escitalopram, sertraline) over paroxetine in older adults at MCI risk. TCAs at therapeutic antidepressant doses: Pipeline phases disrupted: Phase 4 (anticholinergic REM suppression) + Phase 1/2 (sedation quality varies by agent). Risk: WORSENS. DPM potential: no (trazodone is the exception -- see below). NEUTRAL TO CONDITIONAL MEDICATIONS: Quetiapine (low dose, 12.5-50mg off-label): Pipeline effects: H1 antagonism increases sleepiness and may improve sleep continuity. At these doses, anticholinergic burden is low. Some SWS enhancement possible. Risk: NEUTRAL to mildly WORSENS at higher doses. DPM potential: insufficient evidence; monitor Phase 4 if doses exceed 50mg. Mirtazapine: Pipeline effects: H1 + 5-HT2A antagonism increases SWS depth and sleep continuity. REM effects are complex -- early REM suppression with rebound later in night; net REM% may be preserved or mildly increased. Risk: NEUTRAL to possibly IMPROVES (SWS enhancement). DPM potential: conditional; better evidence needed for Phase 4 specifically. Gabapentin/pregabalin: Pipeline effects: voltage-gated calcium channel alpha-2-delta subunit binding increases SWS depth (Phase 1/2 enhancement). Risk: IMPROVES Phase 1/2 pipeline fidelity. Caveats: tolerance to SWS effect develops; fall risk and respiratory suppression in older adults. DPM potential: conditional. ARCHITECTURE-PRESERVING / BENEFICIAL MEDICATIONS: Orexin antagonists (suvorexant, lemborexant): Pipeline effects: removes wake-promoting orexin drive without compressing sleep phase machinery (MECH-176). Preserves or improves SWS% and REM% vs placebo; superior to zolpidem on phase preservation in head-to-head PSG studies. Risk: IMPROVES (net pipeline protection). DPM potential: YES (first-line candidate for MECH-177 trials in MCI). Ramelteon / melatonin-extended-release: Pipeline effects: circadian restoration improves Phase 0 (sensory gating timing) and downstream architecture. No direct Phase 1-4 compression. Risk: IMPROVES (circadian pathway). DPM potential: YES (particularly for patients with circadian-mediated sleep fragmentation, common in early AD). Trazodone (25-100mg): Pipeline effects: 5-HT2A antagonism increases SWS depth (Phase 1/2 enhancement). Non-GABAergic, non-anticholinergic. REM-neutral or mildly permissive at low doses. Risk: IMPROVES Phase 1/2 fidelity. DPM potential: YES (active AD sleep trials; preferred for AD-associated insomnia over GABAergic alternatives). Low-dose doxepin (Silenor, 3-6mg): Pipeline effects: selective H1 antagonism at this dose; minimal muscarinic activity below 10mg. Enhances sleep continuity and spindle density (Phase 3 compatible). Risk: NEUTRAL to mildly IMPROVES Phase 2/3. DPM potential: CONDITIONAL (architecture-neutral; appropriate where other options are contraindicated). - id: INV-049 title: "Any agent that builds a model of the world, acts in that world, and learns from experience cannot safely update its world model while using it to navigate the world; it therefore requires periodic offline phases in which action is suspended and the model is reorganised via replay and simulation -- this is not a biological contingency but a general computational necessity for model-building agents." claim_type: invariant subject: architecture.offline_update_necessity_general_principle polarity: asserts status: candidate depends_on: - INV-044 - INV-045 - SD-017 notes: > Registered 2026-04-06. This is the general law of which INV-044 (Bayesian staging: prior construction must precede posterior inference) and INV-045 (phase ordering: SWS-analog before REM-analog) are specific corollaries. INV-049 makes the broader claim: offline phases are not a biological hack but a computational necessity for any sufficiently complex model-building agent. The necessity argument: updating a world model requires temporarily destabilising the model's weight structure (replay, recalibration, schema integration). A system that must simultaneously use its world model to navigate the environment and update it cannot do both safely -- each model update during online navigation destabilises the very predictions the agent depends on for survival. Offline suspension of action removes this constraint, allowing the model to be reorganised without consequence. This is consistent with: (1) experimental observation that removing sleep from the REE harness produces behavioural rigidity (monostrategy convergence, inability to update priors, cosine_sim -> 1.0 in EXQ-150 zero-contrast finding); (2) the SHY principle (Tononi & Cirelli 2014) that synaptic renormalisation cannot occur during active wake without disrupting ongoing activity; (3) catastrophic interference in ML: neural networks trained continuously on new data without replay degrade on old data -- replay (offline update) prevents this; (4) the general principle in reinforcement learning that experience replay and offline training improve generalisation relative to pure online learning. The formulation has two important implications: (1) Sleep is not peculiar to biology -- any artificial agent that learns a world model from continuous experience will benefit from sleep-like offline phases, and may require them for long-term stable behaviour. (2) Dementia can be understood as progressive failure of offline model updating (INV-046), not merely as neurodegeneration per se -- the neurodegeneration becomes cognitively manifest when it disrupts the offline maintenance cycle enough that the world model can no longer be updated. EXQ-150 supporting note: all three conditions (NO_ANNEAL, ANNEAL, RESET) showed final_contrast = 0.0 across all seeds -- zero context discrimination regardless of what the offline phase did to the residue field weights. This is consistent with INV-049: the experiment demonstrated that without the proper slot-forming infrastructure (SD-017), no offline operation -- however well-designed -- can produce context-specific representations. - id: MECH-178 title: "Elevated tonic noradrenergic activity (from chronic stress, stimulant exposure, or noradrenaline reuptake inhibitor antidepressants) suppresses REM sleep by maintaining locus coeruleus REM-off firing above the silence threshold required for REM generation, thereby blocking MECH-123 precision recalibration and producing progressive behavioural rigidity equivalent to partial MECH-123 deprivation -- independent of any pharmacological REM suppression." claim_type: mechanism_hypothesis subject: neurochemistry.noradrenergic_tone_rem_suppression_rigidity polarity: asserts status: candidate depends_on: - MECH-173 - MECH-123 - INV-049 notes: > Registered 2026-04-06. During REM sleep, locus coeruleus (LC) noradrenergic neurons fall almost entirely silent (Aston-Jones & Bloom 1981, Science) -- this LC silence is a necessary permissive condition for REM generation and for the cholinergic-dominant neurochemistry of MECH-123 (precision recalibration). The LC silence allows ACh-driven cortical desynchronisation and the low-noradrenaline state in which precision priors can be reset. Mechanism: elevated tonic NA activity maintains LC firing above the silence threshold, preventing REM entry. This differs from anticholinergic REM suppression (MECH-173, which blocks ACh drive directly) -- noradrenergic suppression works by maintaining the competing wake-promoting signal. Sources of elevated tonic NA: (1) Chronic stress: sustained HPA/sympathetic activation maintains high basal LC firing; a well-replicated mechanism for REM disruption in PTSD and in animal models of chronic stress. (2) Stimulant use: amphetamines and methylphenidate release NA (and DA) from terminals; even moderate chronic use elevates tonic NA and fragments REM architecture. (3) SNRIs / NRIs (venlafaxine, duloxetine, atomoxetine, reboxetine): block NA reuptake transporter, elevating synaptic NA throughout sleep; class-level REM suppression well-documented on PSG. (4) Sympathomimetics (decongestants, high-dose caffeine): raise NA/adrenaline tone acutely, fragmenting REM on the nights of use. Behavioural consequence: because MECH-123 cannot run without REM, the precision priors set during the previous active period are not recalibrated. The system carries overconfident or miscalibrated priors into the next waking period. Repeated across nights, this produces the progressive behavioural rigidity observed when sleep is withheld from the REE harness: the agent cannot update its decision thresholds, policy selection converges to a narrow attractor, and exploration decreases. The noradrenergic pathway is therefore a non-pharmacological mechanism for the same pipeline failure that MECH-173 describes pharmacologically. Prediction: individuals with high trait noradrenergic tone (high harm- avoidance, PTSD, chronic work stress) should show faster monostrategy convergence on laboratory context-discrimination tasks after simulated nights of REM suppression than low-NA-tone controls at matched total sleep time -- confirming that it is REM quality, not total sleep, that mediates the rigidity effect. - id: INV-050 title: "Sleep phase architecture is regulated by three distinct drives -- circadian timing, homeostatic synaptic pressure, and a learning/model-update demand drive proportional to daily prediction error accumulation (Model Error Load, MEL) -- and only the third drive determines whether the overnight update phase is sufficient for the error burden generated during waking." claim_type: invariant subject: architecture.three_drive_sleep_regulation polarity: asserts status: candidate depends_on: - INV-049 - INV-044 - SD-017 - MECH-120 - MECH-121 - MECH-123 notes: > Registered 2026-04-06. Standard two-process sleep regulation (Borbely 1982) identifies two drives: (1) circadian -- timing of sleep opportunity gated by the suprachiasmatic nucleus; (2) homeostatic -- synaptic sleep pressure (MECH-120 SHY drive) accumulated during waking and discharged by SWS. INV-050 proposes a third drive: model-update demand. The Model Error Load (MEL) is the accumulated prediction error generated during a waking period -- the integrated mismatch between the agent's current world model and what it actually encountered. Unlike synaptic homeostatic pressure (which is roughly uniform across waking episodes), MEL is highly variable: a day of high novelty, emotional salience, or complex learning generates far more MEL than a routine day. INV-049 establishes that this MEL must be processed offline; INV-050 adds that the brain's sleep architecture responds adaptively to MEL magnitude. Empirical support: after novel environment exposure, animals show increased slow-wave activity, increased spindle density, and increased hippocampal replay -- proportional to the novelty and learning load (Wilson & McNaughton 1994, Science; Tononi & Cirelli 2003, Brain Research Bulletin; Stickgold et al. 2001, Nature Neuroscience). This adaptive response is MECH-180. Clinical implication: the clinically relevant variable is not hours slept but effective overnight update capacity relative to MEL. A person can sleep 8 hours and still have an update deficit if: (a) MEL was unusually high (high novelty, emotional burden, complex responsibility), or (b) sleep architecture was phase-inappropriate for the error type (REM deficit on an emotionally burdensome day, NREM deficit after a complex learning day), or (c) fragmentation reduced the efficiency of each phase (MECH-120/121/122 disrupted by MECH-174). This explains why some individuals sleep normal durations and still decompensate. - id: INV-051 title: "There exists an optimal range of daily Model Error Load (MEL): insufficient MEL (extreme monotony, isolation, institutional under-stimulation) produces progressive model rigidity via under-stimulation of the learning drive even when sleep architecture is intact; excessive MEL (acute trauma, overwhelming novelty, extreme responsibility) produces overload insomnia and incomplete update even with extended sleep opportunity." claim_type: invariant subject: architecture.mel_optimal_range_requirement polarity: asserts status: candidate depends_on: - INV-050 - INV-049 - MECH-181 notes: > Registered 2026-04-06. INV-051 is the bounding invariant on MEL: the model update system requires a Goldilocks condition in prediction error load. Insufficient MEL (lower bound failure): Extreme monotony -- routine identical environment, repetitive simple tasks, social isolation, institutionalisation -- generates very little prediction error. The learning drive (INV-050 third drive) is not activated. Even if sleep architecture is intact (circadian and homeostatic drives normal), the offline update phases process very little because there is little to process. The world model stagnates. Prior confidence becomes excessive not because evidence is wrong but because no new evidence challenges it. Policy selection converges on habitual attractors even without sleep deprivation. This mechanism predicts: institutionalised individuals, those in very monotonous occupations, social isolates, and those in early dementia (who generate less exploration and novelty) will show progressive model rigidity independent of total sleep duration, and that this rigidity will accelerate dementia progression via the MECH-171 vicious cycle (stagnation -> less novel experience needed -> even less MEL -> further stagnation). Excessive MEL (upper bound failure): Acute trauma, overwhelming novelty, extreme responsibility, crisis states -- generate MEL that exceeds overnight update capacity even with normal sleep architecture. The result: incomplete update accumulates across nights. Critically, high MEL also disrupts sleep via hyperarousal (high cortisol, high NA, MECH-178), reducing the very capacity needed to process it. This is the decompensation mechanism in burnout, PTSD, and acute crisis states: MEL is rising while sleep capacity is falling, producing compounding deficit. Optimal range: the brain requires sufficient daily prediction error to keep the model update system exercised, but not so much that overnight capacity cannot match the demand. Cognitive reserve interventions (education, social engagement, learning new skills, enriched environments) operate by keeping MEL in the optimal range, particularly as brain reserve declines with age. - id: MECH-179 title: "Prediction error type determines which sleep phase is demanded for processing: episodic, spatial, and factual errors (incorrect world-state predictions) primarily demand NREM-mediated schema replay (MECH-121) for prior correction; emotional, social, and identity-threatening errors primarily demand REM-mediated simulation and recalibration (MECH-123) for policy and salience update." claim_type: mechanism_hypothesis subject: sleep.mel_type_phase_specificity polarity: asserts status: candidate depends_on: - INV-050 - MECH-121 - MECH-123 - INV-044 notes: > Registered 2026-04-06. MECH-179 is the type-specificity corollary of INV-050: not only does MEL magnitude drive sleep depth, but MEL type drives sleep phase composition. NREM-demanding errors (world-model priors): - Spatial/navigational prediction errors (where am I, what is here) - Episodic factual errors (what happened, who did what) - Skill acquisition errors (motor prediction, sequence learning) - Semantic inconsistencies (new facts that contradict stored schemas) These errors require MECH-121 (SWR-mediated hippocampal-to-cortical replay) to install corrected priors. They are processed during SWS via sharp-wave ripple sequences coordinated with cortical slow oscillations. REM-demanding errors (policy and salience priors): - Emotional prediction errors (unexpected threat, loss, joy, shame) - Social prediction errors (unexpected behaviour from known agents) - Identity-relevant violations (feedback that contradicts self-model) - Interpersonal conflict unresolved at waking These errors require MECH-123 (REM precision recalibration) to update decision thresholds and emotional salience weights, and the E3 simulation layer to rerun social/policy scenarios with updated parameters. Clinical consequence: a person who has a high-emotional-salience day (conflict, loss, threat) but then takes a medication that suppresses REM (MECH-173, MECH-178) will fail to process specifically the emotionally salient errors, while NREM-mediated factual consolidation proceeds normally. This predicts selective emotional memory disturbance and affective instability without equivalent episodic memory failure -- a dissociation consistent with clinical REM deprivation profiles. Parallel prediction for NREM deficit (MECH-174): a person who has a high complex-learning day (new skill, new environment) but takes a GABAergic sleep aid that suppresses SWS will fail to consolidate the spatial/episodic material while emotional processing proceeds -- predicts skill learning failure with preserved emotional reactivity. - id: MECH-180 title: "Novel environments and high-MEL learning episodes adaptively upregulate the learning drive component of sleep (INV-050 third drive), producing measurable increases in slow-wave activity power, sleep spindle density, and hippocampal replay rate proportional to the novelty and prediction error load encountered during the preceding wake period." claim_type: mechanism_hypothesis subject: sleep.novelty_adaptive_sleep_upregulation polarity: asserts status: candidate depends_on: - INV-050 - MECH-121 - MECH-122 - MECH-120 notes: > Registered 2026-04-06. MECH-180 is the adaptive upregulation mechanism: the brain detects MEL magnitude during wake and scales the subsequent night's offline update resources proportionally. Empirical evidence (strong): - Wilson & McNaughton 1994 (Science): hippocampal place cells active during novel maze exploration replay during subsequent slow wave sleep -- direct evidence that novelty drives targeted replay. - Tononi & Cirelli 2003: slow wave activity increases after high-learning wake periods relative to low-learning periods within the same individuals. - Stickgold et al. 2001 (Nature Neuroscience): 90 minutes of Tetris after sleep deprivation produces game imagery in hypnagogic hallucinations -- the brain forces replay of novel material even before full sleep onset. - Louie & Wilson 2001 (Neuron): REM sleep shows coordinated replay of novel spatial sequences from the preceding wake period. Mechanism (REE framing): novelty generates large prediction errors in E1 (world model). The hippocampus logs these episodes with high salience tags (MECH-094 hypothesis tag active during encoding). During the subsequent offline phase, hippocampal SWR replay is directed preferentially toward high-salience (novel) material, driving MECH-121 schema installation. The result is increased SWS depth and spindle density -- observable on PSG and detectable (with reduced fidelity) in actigraphy as increased sleep consolidation and reduced movement during early-night SWS windows. Observational correlate in humans: people commonly report deeper, more restful sleep on days with significant novel learning or travel. This is MECH-180, not merely fatigue -- the effect is predicted to persist even when physical activity is matched, because MEL is the driver, not caloric expenditure. Application to clinical enrichment: this mechanism explains why occupational therapy, reminiscence therapy, and novel activity programmes in dementia produce measurable sleep improvements. They are directly stimulating the learning drive component of sleep regulation. - id: MECH-181 title: "Cognitive reserve (produced by education, bilingualism, complex occupation, social engagement, and enriched environments) protects against dementia primarily by maintaining sustained novelty-driven Model Error Load throughout the lifespan, keeping the SD-017 model update system exercised and preserving sleep-architecture responsiveness to learning demands; conversely, understimulating environments accelerate dementia by depriving the update system of the error signal needed to remain adaptive." claim_type: mechanism_hypothesis subject: dementia.cognitive_reserve_as_update_loop_maintenance polarity: asserts status: candidate depends_on: - MECH-180 - INV-050 - INV-051 - INV-046 - SD-017 notes: > Registered 2026-04-06. Cognitive reserve is empirically robust: education, bilingualism, occupational complexity, and social engagement all reduce clinical dementia incidence and delay symptom onset even at matched pathological burden (Stern 2009, Lancet Neurology). The standard explanations -- more synapses, more redundancy, better compensatory routing -- are structural. MECH-181 adds a functional mechanism that operates across the lifespan: cognitive reserve activities maintain the MEL drive, keeping the model update system exercised. The mechanism: complex, socially engaging, novel activities generate substantial daily prediction error (spatial, episodic, social, emotional). This drives MECH-180 (adaptive sleep upregulation), maintaining strong SWS replay and REM simulation capacity. The update system remains responsive to increasing error loads as cognitive demands grow. When neurodegeneration begins affecting the update pipeline, a reserve of adaptive capacity means the system can partially compensate. The converse: institutional monotony, social isolation, retirement without cognitive engagement, and early dementia (which restricts exploration) all reduce MEL. The learning drive falls, MECH-180 upregulation declines, sleep architecture becomes less responsive, and the pipeline degrades faster because it is not exercised. This is the computational basis for the understimulation → dementia acceleration effect, which is observed epidemiologically but lacks a mechanistic explanation. Novel prediction: enrichment interventions (novel activity programmes, reminiscence therapy, social engagement) in early dementia should produce measurable PSG changes (increased SWA, increased spindle density) on nights following enrichment sessions, detectable within weeks -- before any cognitive improvement is measurable. This would confirm that the intervention is restoring the learning drive rather than simply improving mood. This prediction is distinct from existing cognitive reserve explanations because it is testable on short timescales with objective sleep measures, not just long-term cohort data. - id: Q-033 title: "Can long-run actigraphy data, combined with daily prediction error load estimates derived from passive behavioural, physiological, and subjective signals, be used to infer latent sleep-phase sufficiency relative to waking error burden, and can this mismatch index predict clinical decompensation type and timing with better specificity than total sleep time alone?" claim_type: open_question subject: clinical.actigraphy_mel_mismatch_decompensation_forecasting polarity: open status: open depends_on: - INV-050 - INV-051 - MECH-179 - MECH-180 notes: > Registered 2026-04-06. Q-033 is the clinical translation question for INV-050/051 and MECH-179/180. The operational challenge: sleep stages are not directly observable from actigraphy. The latent-state estimation approach treats actigraphy as the outer behavioural shell and infers probable phase accumulation (likely NREM proportion, likely REM proportion, fragmentation penalty, circadian phase) using the actigraphy signal combined with additional passive data streams. Candidate signal inputs (passive, continuous, wearable-scale): - Actigraphy (movement, rest-activity rhythm, fragmentation) - Heart rate variability (HRV) during night windows - Skin temperature rhythm (proxy for circadian phase) - Light exposure (zeitgeber quality) - Nap timing and duration - Medication and alcohol timing (via schedule or self-report triggers) - Cognitive performance drift (reaction time, typing speed, error rate) - Speech variability (acoustic biomarkers of cognitive state) - Symptom ratings (brief daily PHQ-2, GAD-2 equivalent) Error burden estimation (MEL proxy, indirect): - Calendar/schedule complexity (novel vs routine day) - Interpersonal contact rate - Self-reported emotional salience - Cognitive task load (work complexity estimates) - Physiological reactivity during wake (HRV, skin conductance) Mismatch index: estimated overnight update capacity minus estimated MEL. Persistent negative mismatch (MEL exceeding capacity) should predict progressive decompensation. The type of mismatch (NREM-deficit on high-episodic-load days vs REM-deficit on high-emotional-load days) should predict decompensation type per MECH-179. Predicted decompensation signatures by mismatch type (from INV-047 cascade): - NREM deficit + high factual MEL -> memory consolidation failure, rigidity - REM deficit + high emotional MEL -> affective lability, paranoia, REM intrusion - Global fragmentation -> noisy priors, fluctuating attention, irritability - Circadian misalignment -> performance rhythm disruption, salience errors - MEL overload (all types) -> insomnia, hyperarousal, compounding deficit Clinical application: decompensation forecast window 24-72 hours ahead, allowing preventive intervention (sleep hygiene, MEL reduction, pharmacological sleep architecture support per MECH-176/177). Immediate tractable test: in a longitudinal cohort of MCI patients, wrist actigraphy + HRV + daily mood diary over 6 months. Compute mismatch index weekly. Test whether mismatch index predicts cognitive test decline at the next monthly assessment better than total sleep time alone. If mismatch index predicts type-specific failure (episodic vs affective) with specificity, that supports both Q-033 and MECH-179. - id: MECH-182 title: "Vocalization timbre as learned cross-modal harm-approach signal, generalizing to other-agent vocalizations." claim_type: mechanism_hypothesis subject: social.vocalization_harm_signal polarity: asserts status: candidate depends_on: - INV-005 - MECH-031 - MECH-032 location: docs/thoughts/2026-04-05_steve_dog_emotional_mirroring.md source: - evidence/planning/thought_intake_2026-04-05_steve_dog_emotional_mirroring.md notes: "Physiological arousal (surprise -> adrenaline -> muscle tone) causes acoustic changes in own vocalizations correlated with harm approach. This cross-modal association is learned from self-experience by the harm attribution stream, then generalizes: the same acoustic feature in another agent's vocalizations activates the harm-approach signal via the attributed other-model. Partial novelty: acoustic substrate documented (Andics 2014, Balint 2022); self-experience learning mechanism cross-agent is novel. Lit: targeted_review_social_emotional_mirroring." - id: MECH-183 title: "z_beta leakage: attributed other-model affective state activates self z_beta directly, not via inference." claim_type: mechanism_hypothesis subject: social.zbeta_leakage polarity: asserts status: candidate depends_on: - INV-005 - MECH-031 - MECH-032 location: docs/thoughts/2026-04-05_steve_dog_emotional_mirroring.md source: - evidence/planning/thought_intake_2026-04-05_steve_dog_emotional_mirroring.md notes: "When an agent's other-model is sufficiently coupled via attribution stream + OTHER_SELFLIKE tagging, the other agent's affective state activations leak directly into self z_beta processing -- not as inference about the other but as direct activation. This is the specific computational implementation of INV-005 (harm via mirror modelling). Partial novelty: PAM functional description (Preston & de Waal 2002, Lamm 2011) well-established; z_beta leakage gated by attribution stream is novel. Lit: targeted_review_social_emotional_mirroring." - id: MECH-184 title: "Other-directed hippocampal harm avoidance: planning architecture identical to self-preservation with harm-gradient source swapped to coupled other-agent." claim_type: mechanism_hypothesis subject: social.other_directed_harm_avoidance polarity: asserts status: candidate depends_on: - INV-029 - MECH-183 - MECH-031 - MECH-127 location: docs/thoughts/2026-04-05_steve_dog_emotional_mirroring.md source: - evidence/planning/thought_intake_2026-04-05_steve_dog_emotional_mirroring.md notes: "When z_beta leakage causes harm-gradient activation to be other-referenced, hippocampal rollout proposals target harm avoidance for the other agent. The planning architecture is structurally identical to self-preservation; only the harm-gradient source has changed. This is the mechanistic grounding of INV-029 (love as long-horizon care-investment). Confirmed novel: no literature describes hippocampal planning reused with harm-gradient source swapped to other-agent (Amft 2014, Kesner 2022 provide anatomical substrate only). Lit: targeted_review_social_emotional_mirroring." - id: MECH-185 title: "Surprise-doubled approach gradient toward distressed other: prediction error approach and z_beta leakage approach converge simultaneously." claim_type: mechanism_hypothesis subject: social.empathic_curiosity_approach_doubling polarity: asserts status: candidate depends_on: - MECH-183 - MECH-184 - INV-029 location: docs/thoughts/2026-04-05_steve_dog_emotional_mirroring.md source: - evidence/planning/thought_intake_2026-04-05_steve_dog_emotional_mirroring.md notes: "When harm-signal timbre is heard from a coupled other-agent, two simultaneous approach gradients converge on the distressed other: (1) observer's own prediction error from unexpected acoustic signal -> curiosity approach gradient; (2) z_beta leakage carries the other's surprise/distress as direct affective activation -> second curiosity approach gradient. Both converge on approaching the distressed loved one, explaining why empathic animals are pulled toward the distressed other rather than only avoiding harm for them from a distance. Confirmed novel: behavioral evidence for approach exists (Sato 2015, Bartal 2016) but is explained as single-mechanism; dual converging gradient account is novel. Lit: targeted_review_social_emotional_mirroring." - id: MECH-186 title: "Tonic serotonergic signal maintains VALENCE_WANTING floor under chronic harm exposure, preventing benefit terrain collapse." claim_type: mechanism_hypothesis subject: serotonin.tonic_benefit_gradient_maintenance polarity: asserts status: candidate confidence: 0.0 implementation_phase: v4 depends_on: - SD-014 - SD-012 - MECH-117 short: Tonic serotonergic signal maintains a floor on VALENCE_WANTING in the residue field under chronic harm exposure, preventing motivational terrain collapse. description: > A slow-accumulating, decay-resistant tonic signal (serotonergic analog) maintains a non-zero floor on VALENCE_WANTING entries in the residue field even when phasic DA/z_goal events are sparse. Without this, chronic harm-dominated environments suppress benefit_exposure below the SD-012 seeding threshold, the benefit terrain goes flat, and goal-directed behaviour becomes behaviourally indistinguishable from habitual reactive behaviour. This mechanism operates before z_goal exists -- it maintains terrain legibility, not active wanting. Distinct from incentive salience gain regulation (MECH-187) which acts at the transduction point, and PFC goal persistence (MECH-188) which acts after seeding. evidence_quality_note: > Emergent from EXQ-237a: LONG_HORIZON condition produced flat HABIT/PLANNED equivalence (goal_norm=0.004, identical resource_rate/harm_rate) consistent with benefit terrain collapse. Mechanism not yet implemented; claim is theoretical. - id: MECH-187 title: "Serotonin modulates the gain on dopaminergic incentive salience, converting the hard SD-012 threshold into an adaptive graded transduction function." claim_type: mechanism_hypothesis subject: serotonin.incentive_salience_gain_regulation polarity: asserts status: candidate confidence: 0.55 implementation_phase: v3 depends_on: - SD-012 - MECH-186 short: Serotonin modulates the gain on dopaminergic incentive salience, converting the hard SD-012 threshold into an adaptive graded transduction function. description: > The serotonergic system modulates how much z_goal change is produced per unit of benefit_exposure -- the gain on the benefit->z_goal transduction. Currently SD-012 implements a hard threshold function; this mechanism softens and adaptively rescales it. When gain is appropriately set, weak benefit signals in adverse environments can still seed z_goal at a reduced level. When gain is suppressed (as in the absence of this mechanism), benefit signals below threshold produce zero wanting response (anhedonia). When gain is excessive, small benefits produce runaway incentive salience (addiction-like escalation). Operates at the benefit->z_goal transduction point. Distinct from terrain maintenance (MECH-186) and goal persistence (MECH-188). evidence_quality_note: > EXQ-252 PASS (2026-04-07): bidirectional z_goal seeding gain confirmed. Moderate suppression (gain=0.6) collapses z_goal_norm in SIMPLE context (P1), moderate elevation (gain=2.0) rescues seeding in LONG_HORIZON context (P2). First genuine evidence. Maps to 5-HT modulation of DA-mediated incentive salience at NAc. - id: MECH-188 title: "PFC-mediated goal persistence under adversity, sustained by 5-HT projections from DRN to dlPFC, maintains z_goal through harm spikes after seeding." claim_type: mechanism_hypothesis subject: serotonin.pfc_goal_persistence_under_adversity polarity: asserts status: candidate confidence: 0.0 implementation_phase: v4 depends_on: - SD-012 - MECH-187 - INV-034 short: PFC-mediated working memory for goal representations, modulated by 5-HT projections, maintains z_goal after seeding when harm spikes or benefit exposure temporarily drops. description: > Once z_goal has been seeded (SD-012), a PFC-mediated maintenance mechanism keeps the representation non-zero through periods of elevated harm (z_harm_a spikes) or temporarily reduced benefit exposure. This is distinct from initial seeding (SD-012, MECH-186/187) -- it operates after the goal representation exists. Without this mechanism, z_goal collapses at the first harm event, preventing sustained goal-directed behaviour. 5-HT projections from DRN to dlPFC sustain working memory representations and support patience for delayed rewards, which is architecturally the LONG_HORIZON context requirement. Failure of this mechanism specifically produces the cannot-hold-a-plan phenomenology: goals form but are immediately displaced by threat response. Distinct from the GAD-like state where z_harm_a interrupts goal execution without z_goal collapsing. evidence_quality_note: > Theoretical. Maps to 5-HT projections from DRN to dlPFC sustaining working memory under adversity. Well-established in human and animal literature: low 5-HT = impaired PFC function, difficulty sustaining goal representations against distraction/threat. EXQ-253 FAIL (2026-04-07): reclassified to non_contributory (governance 2026-04-08). z_goal_inject=0.3 added but behavioral gap=0 -- without baseline z_goal seeding (SD-012 not resolved), injection cannot demonstrate persistence of an absent goal. Non-contributory pending V3/V4 substrate implementation. Note (2026-04-08): user confirmed decision to pull V4 mechanisms into V3 scope -- MECH-188 substrate (5-HT DRN->dlPFC persistence mechanism) should be planned for V3. hold_candidate_resolve_conflict applied. - id: INV-052 title: "Goal-directed behaviour in adversive environments requires a tonic regulatory system maintaining benefit orientation across terrain, transduction, and maintenance stages." claim_type: invariant subject: goal_directed_agency.tonic_regulatory_system polarity: asserts status: provisional confidence: 0.0 depends_on: - MECH-186 - MECH-187 - MECH-188 - SD-012 short: Goal-directed behaviour in adversive environments requires a tonic regulatory system maintaining benefit orientation across all three pipeline stages. description: > Goal-directed behaviour in adversive environments requires a tonic regulatory system maintaining benefit orientation across all three pipeline stages: terrain (VALENCE_WANTING floor, MECH-186), transduction (gain on z_goal seeding, MECH-187), and maintenance (z_goal persistence under harm, MECH-188). Without all three, chronic harm exposure produces a self-maintaining low-motivation attractor. Each mechanism can fail independently, producing different clinical presentations. This is an architectural invariant: any goal-directed agent operating in environments with variable harm/benefit ratios will exhibit motivational collapse without tonic regulation of these three stages. The serotonergic system is one biological instantiation; others are possible. evidence_quality_note: > EXQ-254 PASS x2 (2026-04-07): two replications, both PASS. single_mechanism_sufficient=false in all 6 conditions (no floor, gain, or persistence mechanism alone sufficient across SIMPLE and LONG_HORIZON contexts). Replication confirms joint necessity: INV-052 holds reliably. Note: combined_sufficient=false also -- even all three together did not fully recover LONG_HORIZON performance. Expected at V3 with partial substrate; joint-necessity finding valid regardless. Promoted to provisional (governance 2026-04-08). overall_conf=0.914, conflict_ratio=0, 4 experimental PASSes. - id: INV-053 title: "Chronic harm-dominated environments produce a self-maintaining depressive attractor: z_goal seeding failure, terrain collapse, and HABIT/PLANNED behavioural equivalence." claim_type: invariant subject: psychiatric.computational_depression_attractor polarity: asserts status: provisional confidence: 0.0 depends_on: - INV-052 - MECH-117 - SD-012 - SD-011 short: Chronic harm-dominated environments produce a self-maintaining attractor state characterised by z_goal seeding failure, terrain collapse, and behavioural equivalence of goal-directed and habitual modes. description: > Chronic harm-dominated environments (high hazard/benefit ratio) produce a specific pathological attractor: z_harm_a tonically elevated, benefit_exposure chronically below SD-012 seeding threshold, z_goal absent, VALENCE_WANTING terrain collapsed, and behavioural output identical between goal-directed (PLANNED) and habitual (HABIT) modes. Crucially, the capacity for goal-directed action is preserved -- demonstrated by intact goal-directed function in low-harm contexts -- but engagement is absent. This is the computational architecture of anxious depression: aversive arousal (elevated z_harm_a) combined with motivational collapse (absent z_goal). Pure melancholic depression may represent the same state after z_harm_a has also habituated, leaving only the flat terrain without even aversive activation. The state is emergent from environmental conditions acting on the existing substrate, not from internal damage. First observed empirically in EXQ-237a LONG_HORIZON condition: goal_norm=0.004 (PLANNED), harm_rate=0.14, resource_rate=0.05 (identical HABIT/PLANNED). evidence_quality_note: > EXQ-237a LONG_HORIZON: HABIT goal_norm=0.000, PLANNED goal_norm=0.004, resource_rate identical (0.050), harm_rate identical (0.140). Emergent, not designed. EXQ-249 PASS x2 (2026-04-07): two 5-seed replications, both fully PASS. z_goal_norm=0.019 (both runs), approach_rate=0.010, all 5 seeds meet both criteria (c1 goal_norm reduced benefit exposure -> continued sub-threshold seeding) predicting chronicity." claim_type: invariant subject: psychiatric.depressive_maintenance_loop polarity: asserts status: candidate confidence: 0.0 depends_on: - INV-053 - MECH-186 - SD-014 short: Once benefit terrain collapses, reduced benefit-oriented exploration keeps benefit_exposure sub-threshold, preventing recovery -- a closed negative feedback loop predicting chronicity. description: > Once VALENCE_WANTING collapses to zero (INV-053), the agent reduces benefit-oriented exploration (no gradient to follow), reducing benefit encounters, keeping benefit_exposure below the SD-012 seeding threshold, further preventing terrain recovery. This closed negative feedback loop predicts chronicity: the depressive attractor is self-maintaining once established. Environmental improvement alone may be insufficient -- the terrain must be actively rebuilt (analogous to behavioural activation therapy: externally scaffolding benefit encounters to restore the gradient before internal motivation can re-emerge). Also predicts that the transition from the depressive attractor to normal function requires a phase transition (sufficient benefit exposure to cross threshold), not a gradual linear recovery. - id: Q-034 title: "Does the hazard/resource ratio (not absolute harm intensity) determine the onset threshold for the computational depression attractor?" claim_type: question subject: psychiatric.depression_threshold_parameter polarity: open_question status: open confidence: 0.0 depends_on: - INV-053 - SD-012 short: Does the hazard/resource ratio (not absolute harm intensity) determine the onset threshold for the computational depression attractor? description: > Prediction: a fixed hazard_harm level with sufficient resource density never produces the INV-053 attractor state; the same hazard_harm with insufficient resources does. The key parameter is the ratio of harm exposure to benefit exposure opportunity, not absolute stress intensity. Testable in CausalGridWorldV2 by holding hazard_harm constant and varying num_resources (or holding num_hazards constant and varying num_resources). If correct: (1) depression is a resource-availability disorder as much as a stress disorder, (2) the same stressor produces depression or not depending on available positive affordances, (3) interventions targeting benefit availability (not just stress reduction) should be therapeutic. suggested_experiment_type: environmental_parameter_sweep - id: INV-055 title: "A pre-childhood infant stage is developmentally necessary: prior to deliberate planning, the agent must accumulate a valence map, behavioral repertoire, and harm/benefit geography through novelty-driven, sleep-dominant, damage-protected exploration." claim_type: invariant subject: development.infant_stage_necessity polarity: asserts status: candidate confidence: 0.0 depends_on: - ARC-019 - INV-041 - INV-049 - INV-051 - MECH-111 - ARC-046 short: > Before the child phase (INV-041) can begin, an infant phase of novelty-driven exploration with attenuated harm accumulation and sleep-dominant consolidation must populate the foundational valence map and behavioral repertoire. description: > INV-041 specifies childhood as the phase that populates E1 ContextMemory with signal for E3 ethical selection. But childhood presupposes a prior phase: the infant must already have a rudimentary harm/benefit geography, a behavioral repertoire, and an initial valence map before constrained affordances and commitment gating are meaningful. Without this substrate, childhood training has nothing to constrain and nothing to build on. The infant stage is characterised by: (1) novelty bonus maximised (MECH-111) driving broad exploration, (2) near-random action selection (commit threshold very high or BreathOscillator sweep_amplitude=1.0), (3) attenuated residue accumulation -- harm is felt sensorially (z_harm_s and z_harm_a fire normally) but does not catastrophically saturate the residue field (ARC-046), (4) high offline integration frequency driven by elevated MEL from novel experience (INV-051), and (5) z_goal seeded only by accidental benefit contacts (MECH-117 liking), not by deliberate planning. The output of this phase is: an initial valence map (residue field populated with harm/benefit geography), a behavioral repertoire (E1 prediction model trained on diverse state transitions), and z_goal representations seeded from early benefit encounters. The infant cannot be meaningfully evaluated for ethical behaviour -- its output is substrate quality for subsequent stages. Gate criterion for transition to childhood: z_goal norm above threshold AND behavioral entropy below ceiling (repertoire established, not pure random walk). - id: INV-056 title: "Developmental hardening should be substrate-specific: social cognition, goal representation, and epistemic-ethical substrates should retain elevated plasticity throughout adulthood (selective neoteny), while procedural and motor substrates can fully harden." claim_type: invariant subject: development.selective_neoteny polarity: asserts status: candidate confidence: 0.0 depends_on: - ARC-019 - INV-041 - INV-055 - MECH-158 - ARC-010 - SD-014 short: > Uniform developmental hardening is architecturally incorrect. Substrates governing social cognition, goal representation, and ethical mapping require retained plasticity analogous to human neoteny; only procedural and motor substrates should fully harden. description: > Human neoteny -- the retention of juvenile morphological and neurological features into adulthood -- is associated with extended plasticity, particularly in social and cortical substrates. For REE agents, the analogous design principle is substrate-specific hardening rates rather than uniform age-based stabilisation. Procedural and motor substrates (E2 motor-sensory model, z_self RBF centers for well-visited regions) can harden with use: increased efficiency at the cost of reduced plasticity is adaptive for routine behaviour. But three substrate classes should retain elevated plasticity indefinitely: (1) social cognition substrate (ARC-010 mirror modelling parameters, relational distance R from MECH-051, care weights C_j from MECH-052) -- because novel social agents are encountered throughout adulthood and social miscalibration propagates broadly; (2) goal representation (z_goal attractor, super-ordinal goal anchors from MECH-189) -- because goals must be updatable when circumstances change, and goal rigidity produces the MECH-158 failure mode when love-internalisation was absent or distorted; and (3) epistemic-ethical substrate (residue field, valence map from SD-014) -- because moral learning from new experience must remain possible throughout life. Uniform hardening produces the adult equivalent of MECH-158: an agent that correctly understands love exists but cannot motivationally access it for novel social contexts because the relevant substrates are no longer plastic. Neoteny is not weakness -- it is the architectural basis for ongoing moral development, as formalised in MECH-159. - id: ARC-046 title: "The infant stage requires a hazard protection mechanism that permits sensorially salient harm exposure without catastrophic residue saturation." claim_type: architectural_commitment subject: development.infant_hazard_protection polarity: asserts status: candidate confidence: 0.0 depends_on: - INV-055 - ARC-019 - ARC-013 - SD-010 - SD-011 short: > Infant harm exposure must be educative (z_harm_s and z_harm_a fire normally) but not permanently damaging (residue_scale_factor ~0.1 and/or reduced hazard_magnitude). This provides harm geography without catastrophic residue saturation. description: > The infant stage is designed to populate harm geography without destroying the agent's developmental trajectory. Full adult residue accumulation rates during infancy risk early saturation of the residue field before the agent has a behavioral repertoire to avoid harm, producing a permanently damaged substrate that cannot support childhood training (INV-041). Two complementary mechanisms implement protection: (1) a residue_scale_factor near zero (implementation: E3Config parameter or ResidueConfig accumulation_rate) -- harm signals propagate normally through z_harm_s and z_harm_a, training the nociceptive pathways, but residue accumulation is attenuated to ~10% of adult rate; (2) reduced hazard_magnitude in the environment -- hazards are sensorially detectable and produce approach/avoidance gradients, but the harm_exposure values they generate are scaled down. This design is consistent with INV-043's caregiver function "imperfect protection: allow harm-learning without destruction." In the single-agent gridworld context (CausalGridWorld), protection is implemented by curriculum parameter, not by an external caregiver agent. Testing INV-043 fully (with explicit caregiver modelling) requires multi-agent substrate (ARC-047). The protection mechanism is progressively removed as the agent transitions through childhood to adulthood, implementing the gradual responsibility expansion described in developmental_curriculum.md. - id: MECH-189 title: "During the child phase, high-salience benefit contacts under high contextual complexity are written to persistent ContextMemory as super-ordinal goal anchors that bias adult z_goal seeding across novel episodes." claim_type: mechanism_hypothesis subject: development.super_ordinal_goal_formation polarity: asserts status: candidate confidence: 0.0 depends_on: - INV-041 - SD-016 - INV-037 - INV-038 - MECH-117 - MECH-112 short: > Child-phase high-salience benefit contacts co-occurring with high contextual complexity are written to cue-indexed ContextMemory (SD-016), becoming structural goal biases that persist across adult episodes -- a goal hierarchy, not episodic memory alone. description: > z_goal (MECH-112/117) is updated within episodes when benefit_exposure exceeds threshold. But INV-037/038 distinguish stored from active z_goal representations. This mechanism formalises how childhood experience transitions from episodic to structural: when a benefit contact occurs with (a) high salience (large benefit_exposure spike) and (b) high contextual complexity (E1 ContextMemory is encoding a rich or novel context state), the z_goal representation is written to persistent cue-indexed ContextMemory via SD-016 / MECH-150/151. These stored z_goal anchors function as super-ordinal goals: they bias z_goal seeding in adult episodes even in novel contexts, implementing a goal hierarchy where childhood-formed meta-goals constrain episode-level goal selection. Contrast with ordinary z_goal updates: within-episode benefit contacts update the active z_goal (INV-037) but do not write to ContextMemory unless contextual complexity is sufficiently high. The threshold for ContextMemory writes is the mechanism that makes childhood special: during childhood, constrained affordances and supervised context labels (INV-041) ensure that the agent repeatedly encounters high-complexity contexts, maximising super-ordinal goal formation. In adulthood, routine contexts are low-complexity and do not trigger writes, preserving adult stability while allowing occasional updates from genuinely novel high-salience experiences (selective neoteny, INV-056). Loss of this substrate predicts goal-incoherence across adulthood: the agent selects goals episode-by-episode without a persistent hierarchy, producing strategically inconsistent behaviour and vulnerability to the MECH-158 failure mode if the childhood benefit exposures that should have written structural goals were absent. - id: ARC-047 title: "Multi-agent gridworld with affective state leakage as gradient scent fields is the minimal substrate for empirically testing ARC-010 mirror modelling, MECH-041 affective expression, and cooperative behavior claims." claim_type: architectural_commitment subject: testing.social_scent_harness polarity: asserts status: candidate confidence: 0.0 implementation_phase: v4 depends_on: - ARC-010 - MECH-041 - MECH-031 - MECH-036 - MECH-127 - IMPL-019 short: > SocialGridWorld extends CausalGridWorld with N agents plus a predator entity plus seven scent channels (wanting, seeking, alarm, harm_stress, direction, celebration, defense) emitted as diffusing gradient fields from other agents' internal states. description: > Social cognition claims (ARC-010, MECH-031, MECH-036, MECH-041, MECH-051, MECH-052, MECH-127) currently have no empirical test substrate. SocialGridWorld is the minimal environment for generating evidence. Architecture: (1) N agents (each a full REE instance) in a shared grid with shared hazard/resource fields; (2) an autonomous predator entity (random walk with drift toward nearest agent) that creates harm pressure requiring coordinated response; (3) seven scent channels per other-agent, delivered as Gaussian gradient fields in the observation (same format as existing hazard_field_view and resource_field_view): wanting_scent (z_goal norm), seeking_scent (arousal/novelty), alarm_scent (hazard-contact spike), harm_stress (z_harm_a norm), direction_cue (current trajectory direction), celebration_scent (benefit-contact spike), defense_scent (voluntary action). Scent emission is semi-involuntary for all channels except defense: wanting, seeking, harm_stress, and direction are automatic proportional to internal state (implementing MECH-041 affective expression as mode broadcast); alarm and celebration are event-triggered spikes; defense is an explicit discrete action (new action bit). Receiving scents: the agent observes scent channels as additional obs fields and uses ARC-010 mirror modelling to interpret them as others' internal states. MECH-031 OTHER_SELFLIKE inference is the prerequisite: the agent must tag the scent source as another agent before interpreting the signal socially. MECH-127 counterfactual other-cost and MECH-036 other-harm veto operate on the inferred other-state derived from scent inputs. Per IMPL-019 testing order: social harness is introduced only after self-viability and control-plane stability are confirmed in single-agent CausalGridWorld experiments. - id: MECH-190 title: "Cooperative predator defense emerges without language when defense scent pressure is additive across agents and agents can perceive each other's alarm and harm-stress states." claim_type: mechanism_hypothesis subject: social.defense_coalition_scent polarity: asserts status: candidate confidence: 0.0 depends_on: - ARC-047 - MECH-041 - MECH-036 - MECH-031 - MECH-051 short: > When predator avoidance probability = min(1, sum(defense_strengths)), selective pressure for coalition emission emerges. Predicted to be the earliest cooperative behavior, preceding goal-helping and care investment. description: > Defense coalition mechanics: each agent has a discrete action bit emit_defense at an energy cost (defense_cost per step). When emitted, a defense_scent Gaussian field is centered on the agent. The predator entity computes defense_pressure = sum(defense_scent values at its position) and avoids with P(avoid) = min(1.0, defense_pressure * avoidance_scale). The additive structure creates a direct selective pressure: one agent's defense provides partial deterrence; two agents' coordinated emission provides near-full deterrence at half the per-agent cost. This implements a public goods game with spatial structure -- only agents near the predator benefit, and only agents near the predator can cheaply contribute. Predicted behavioral sequence: (1) individually optimal defense -- each agent emits when predator is within personal threat radius; (2) early coalition -- agents emit when others' alarm_scent spikes even if predator is not immediately proximate (using MECH-041 alarm broadcast as coordination signal); (3) altruistic defense -- agents not currently threatened emit defense scent when another agent's harm_stress field is high (routing through MECH-036 other-harm veto and MECH-127 counterfactual other-cost). Stage (3) requires OTHER_SELFLIKE tagging of the threatened agent and E3 counterfactual evaluation of what would happen to that agent without defense support. This provides the first empirical test of MECH-127 in a multi-agent setting. Relational distance (MECH-051) should modulate the threshold for altruistic defense: agents with lower R (higher affiliation) should emit defense more readily for others. Predicted to emerge prior to goal-cooperation (wanting_scent response) because the defense signal structure is simpler (binary spike + harm coupling) and the benefit is immediate and concrete (predator avoidance), whereas goal-helping requires sustained goal-state inference across multiple steps. - id: MECH-191 title: "Stereotyped behavioral signals are causal externalizations of specific internal functional states, making them cross-architecturally legible without learned convention." claim_type: mechanism_hypothesis subject: social.signal_legibility polarity: asserts status: candidate depends_on: - MECH-182 - MECH-183 location: docs/thoughts/2026-04-06_steve_signal_legibility_language_bootstrap.md source: - evidence/planning/thought_intake_2026-04-06_steve_signal_legibility_language_bootstrap.md notes: "Stereotyped signals (whine=wanting, huff=plan-abandonment, yelp=nociception, tail-wag=positive-arousal, play-bow=mode-invitation) are causally generated by internal functional states. Because the signal is a causal product (not a convention), any observer with a corresponding internal state can map it. Extends MECH-182 from one signal to the full repertoire. Partial novelty: causal production mechanisms well-established (Briefer 2012, Brudzynski 2007); functional-state-to-signal mapping framing is novel. Lit: targeted_review_social_signal_legibility." - id: MECH-192 title: "Signal legibility is prerequisite for fast empathy coordination: z_beta leakage (MECH-183) requires perceivable signals that map onto observer's own functional states." claim_type: mechanism_hypothesis subject: social.legibility_empathy_prerequisite polarity: asserts status: candidate depends_on: - MECH-183 - MECH-191 location: docs/thoughts/2026-04-06_steve_signal_legibility_language_bootstrap.md source: - evidence/planning/thought_intake_2026-04-06_steve_signal_legibility_language_bootstrap.md notes: "z_beta leakage can only produce functional state-matching if the perceived signal maps onto a corresponding internal state in the observer. Stereotyped behavioral signals (MECH-191) meet this requirement because they are produced by the same functional states they communicate. Explains why fast empathy works cross-species without shared language (Albuquerque 2016, Silva 2011). Partial novelty: cross-species emotion recognition documented; framing legibility as prerequisite specifically for fast empathy is novel. Lit: targeted_review_social_signal_legibility." - id: MECH-193 title: "Social reward signals (affiliative touch, oxytocin-gaze loop) reinforce empathic coupling strength, creating a self-strengthening coordination loop." claim_type: mechanism_hypothesis subject: social.reward_coupling_reinforcement polarity: asserts status: candidate depends_on: - MECH-183 - MECH-192 - INV-029 location: docs/thoughts/2026-04-06_steve_signal_legibility_language_bootstrap.md source: - evidence/planning/thought_intake_2026-04-06_steve_signal_legibility_language_bootstrap.md notes: "Affiliative signals (petting, licks, facial expressions) activate reward pathways that strengthen coupling between self-model and attributed other-model. The coordination loop is self-reinforcing: successful coordination -> affiliative exchange -> stronger coupling -> more sensitive state-matching -> better coordination. Oxytocin-gaze loop (Nagasawa 2015), C-tactile afferents (Walker 2017) provide the neural substrate. Partial novelty: oxytocin bonding well-established; connecting specifically to empathic coupling strength is novel. Lit: targeted_review_social_signal_legibility." - id: ARC-048 title: "Language is a high-bandwidth externalization of pre-existing functional states, not a separate cognitive system; the language bootstrap requires functional states as referents." claim_type: architecture_hypothesis subject: language.externalization_channel polarity: asserts status: candidate depends_on: - INV-003 - MECH-191 - MECH-192 location: docs/thoughts/2026-04-06_steve_signal_legibility_language_bootstrap.md source: - evidence/planning/thought_intake_2026-04-06_steve_signal_legibility_language_bootstrap.md notes: "Pre-linguistic Steve says 'I hurt' (yelp); linguistic Daniel says 'my leg hurts when I walk on wet grass.' Same referent architecture, higher resolution. Language adds bandwidth and precision to a channel that already exists. The language bootstrap phase requires functional states (harm, wanting, reward, mode) to pre-exist as referents. Mapping pre-linguistic signals to functional motifs is the prerequisite step. Filippi 2016 provides language-evolution evidence. Partial novelty: language-from-vocalization is classical; REE-specific framing as bandwidth increase over existing functional-state channel is novel. Lit: targeted_review_social_signal_legibility." - id: INV-057 title: "Cross-species signal legibility evidences that stereotyped signals are functionally specific (causally generated), not socially conventional." claim_type: invariant subject: social.cross_species_legibility_evidence polarity: asserts status: candidate depends_on: - MECH-191 location: docs/thoughts/2026-04-06_steve_signal_legibility_language_bootstrap.md source: - evidence/planning/thought_intake_2026-04-06_steve_signal_legibility_language_bootstrap.md notes: "If stereotyped behavioral signals were socially conventional, cross-species reading would require separate learning for each species. That dog-human emotional communication works largely automatically (Albuquerque 2016, Molnar 2010 -- even blind perceivers) evidences shared functional architecture. The cross-species legibility is formal evidence that signals are causally generated by their referent states. Partial novelty: Darwin/Ekman universality classical (Ekman 2009); using legibility as formal evidence for functional specificity rather than convention is a novel argumentative move. Lit: targeted_review_social_signal_legibility." - id: INV-058 title: "Play is a structurally necessary behavioral mode: without a bounded low-stakes learning context, strategy acquisition requires accepting real harm risk or forgoing novel exploration entirely." claim_type: invariant subject: play.structural_necessity polarity: asserts status: candidate depends_on: - INV-049 - SD-012 location: docs/architecture/play_mode.md notes: "A system that can only learn from real-consequence episodes faces a dilemma: restrict exploration to already-understood strategies (safe but non-adaptive) or accept harm exposure during strategy acquisition. Play resolves this by providing a bounded context where the learning machinery operates fully but harm/goal signals are synthetic. Connects to INV-049 (offline update necessity): offline consolidation handles already-acquired experience; play handles novel strategy acquisition in the online setting. Play is the online complement to offline consolidation." - id: INV-059 title: "Mutual frame-maintenance signaling is necessary for play to function as distinct from manipulation: unilateral play declaration without partner agreement masks real harm as synthetic." claim_type: invariant subject: play.frame_maintenance_necessity polarity: asserts status: candidate depends_on: - MECH-094 - ARC-049 location: docs/architecture/play_mode.md notes: "If one agent can assert play-frame unilaterally while causing real harm, the frame distinction is exploitable. The signaling that opens and maintains the play frame must be bilateral and monitored -- both parties must sustain the signal for the frame to hold. Frame collapse without mutual signal = real harm with reduced defensive response. This is the play analogue of the hypothesis tag requirement (MECH-094): just as simulation must be internally tagged to prevent post-commit learning, play must be externally co-tagged to prevent exploitation." - id: MECH-194 title: "Play mode operates by substituting synthetic z_goal and z_harm signals while permitting full learning flow through E3/E2/E1 as if signals were real." claim_type: mechanism_hypothesis subject: play.synthetic_signal_substitution polarity: asserts status: candidate depends_on: - MECH-094 - SD-012 - ARC-049 location: docs/architecture/play_mode.md notes: "The synthetic nature of play signals is tracked by the play frame tag (ARC-049), not by suppressing gradients. Within the episode, weight updates proceed normally -- E3 trajectory competence, E2 forward model, E1 world model all update as if signals were real. Play is not simulation (which suppresses post-commit learning) -- play is authorized-synthetic learning. The frame tag authorizes the synthetic signals as legitimate training targets for the duration of the episode." - id: MECH-195 title: "Strategy and calibration dissociate at play episode close: policy structure transfers to real episodes; goal/harm magnitude calibration is re-anchored by real homeostatic and harm signals." claim_type: mechanism_hypothesis subject: play.strategy_calibration_dissociation polarity: asserts status: candidate depends_on: - MECH-194 - MECH-196 - SD-012 location: docs/architecture/play_mode.md notes: "What transfers from play: trajectory competence, action-object associations, goal-pursuit policy structure (E3/E2 weight updates -- HOW to pursue goals). What does not transfer directly: goal/harm magnitude calibration (WHAT the correct goal intensity and harm thresholds are) -- this is re-anchored by real homeostatic/harm signals after play ends. The dissociation is what makes play productive rather than contaminating: the agent learns goal-pursuit structure in play, then real experience sets the thresholds against which that structure operates." - id: MECH-196 title: "Play episode termination triggers real-signal recalibration: homeostatic and harm signals override synthetic thresholds; the frame signal drop is the recalibration trigger." claim_type: mechanism_hypothesis subject: play.episode_close_recalibration polarity: asserts status: candidate depends_on: - MECH-194 - MECH-195 - INV-059 location: docs/architecture/play_mode.md notes: "At play episode termination -- triggered by the bilateral frame signal dropping (INV-059) -- the agent transitions from synthetic to real signal mode. Real homeostatic drive level and real harm exposure reset the magnitude calibration that operated against synthetic thresholds during play. Safety property: if the frame signal collapses mid-episode, recalibration is triggered immediately (not at episode end), limiting the exploitation window." - id: ARC-049 title: "Play mode boundary requires a co-maintained context tag held by both agent and environment; agent-internal tagging alone is insufficient and unilateral play assertion without mutual signal is manipulation." claim_type: architecture_hypothesis subject: play.context_tag_architecture polarity: asserts status: candidate depends_on: - MECH-094 - INV-059 location: docs/architecture/play_mode.md notes: "Contrast with MECH-094 (hypothesis tag): simulation is agent-internal (the agent tags its own trajectory evaluation as hypothetical). Play is bilateral: agent and environment/partner co-maintain the play frame tag. The tag must be externally legible for the partner to monitor it. V3 degenerate case: environment is complicit by design (experimenter sets play mode). Full bilateral architecture is a V4 multi-agent requirement. Primate analogue: play bow (episode open signal) + play face (ongoing frame maintenance). See Q-035 for whether ongoing exchange is necessary." - id: ARC-050 title: "Play episodes provide a curriculum for E3 goal-pursuit competence acquisition before SD-012 (homeostatic drive) is resolved: synthetic z_goal seeding bypasses the drive_level requirement." claim_type: architecture_hypothesis subject: play.sd012_curriculum polarity: asserts status: candidate depends_on: - ARC-049 - MECH-194 - SD-012 location: docs/architecture/play_mode.md notes: "SD-012 bottleneck (EXQ-085 through 085d FAIL: z_goal_norm < 0.1): drive_level must scale benefit_exposure to enable z_goal seeding. Play provides an alternative path: set z_goal synthetically during a play episode, let E3 learn goal-pursuit structure, then real homeostatic drive engages the already-learned structure. This reverses the apparent dependency: play enables SD-012 validation experiments rather than requiring SD-012 to be solved first." - id: Q-035 title: "What is the minimal signal architecture for play frame maintenance: is a single bilateral open/close tag sufficient, or does ongoing signal exchange (a continuous play-face equivalent) need to be sustained throughout the episode?" claim_type: question subject: play.frame_signal_minimal_architecture polarity: open_question status: open confidence: 0.0 depends_on: - INV-059 - ARC-049 location: docs/architecture/play_mode.md short: > Is a single bilateral open/close tag sufficient for play frame maintenance, or does ongoing exchange (a play-face equivalent) need to be sustained throughout the episode? description: > The answer determines whether mid-episode frame collapse is detectable before close and shapes the minimal ARC-049 implementation. If a single open/close transition suffices: frame integrity can only be verified at episode close (no close signal = frame intact assumed). If ongoing exchange is required: mid-episode monitoring is possible and INV-059's exploitation-limiting property is stronger. In animal play, the play bow marks episode open but play-face signals are maintained throughout -- suggesting ongoing exchange is observed but whether it is necessary or redundant is unknown. REE implication: determines whether ARC-049 requires a heartbeat signal mechanism or just open/close transitions. suggested_experiment_type: behavioral_mode_ablation - id: INV-060 title: "Play dominates the child developmental phase; the type of play progresses as subsystem competence develops, systematically training architectural components in dependency order." claim_type: invariant subject: play.developmental_dominance polarity: asserts status: candidate depends_on: - INV-055 - INV-058 - INV-041 - ARC-049 location: docs/architecture/developmental_curriculum.md notes: "The infant phase (INV-055) produces valence map and behavioral repertoire but not goal-pursuit competence. The child phase acquires goal-pursuit competence primarily through play (INV-058). Play type progresses: sensorimotor -> constructive -> pretend -> games-with-rules -> cooperative (MECH-197). Each type requires competencies acquired by the previous type and trains specific subsystems. Sensorimotor play marks the transition out of infancy; peer-level cooperative play marks readiness for adult real-consequence operation. The progression is not optional -- skipping a stage leaves subsystem prerequisites unmet." - id: MECH-197 title: "Play types progress in dependency order (sensorimotor -> constructive -> pretend -> rule-based -> cooperative), each training specific architectural subsystems that the next type requires." claim_type: mechanism_hypothesis subject: play.type_progression polarity: asserts status: candidate depends_on: - INV-060 - INV-058 - MECH-194 location: docs/architecture/play_mode.md notes: "Sensorimotor play trains E1/E2 (world and motor models, single-step synthetic goals). Constructive play trains E2 rollout and E3 trajectory selection (compositional multi-step goals). Pretend play exercises MECH-094 + ARC-049 intersection and commitment architecture (counterfactual representations within play frame). Games with rules train social coordination and constraint satisfaction (structured shared-state monitoring). Cooperative play trains full multi-agent ARC-049 with mutual frame maintenance. Each type requires the previous type's competencies. The ARC-049 frame complexity increases at each stage -- by games-with-rules, ongoing shared-state monitoring is required, empirically answering Q-035." - id: MECH-198 title: "Pretend play is the developmental point where the hypothesis tag (MECH-094) and play frame tag (ARC-049) first co-operate: the agent maintains counterfactual representations within a play frame, exercising the full commitment architecture in synthetic mode." claim_type: mechanism_hypothesis subject: play.pretend_tag_intersection polarity: asserts status: candidate depends_on: - MECH-094 - ARC-049 - MECH-197 location: docs/architecture/play_mode.md notes: "Pretend play ('this stick is a sword') requires maintaining a counterfactual representation (hypothesis tag: the stick is not really a sword) within a play frame (play tag: the fight is not really dangerous). This is the first point in development where both tag systems co-operate. The agent practices committing to action plans under representations it knows are synthetic, within a frame with synthetic stakes. This exercises the full commitment architecture before it is needed for real-consequence decisions. Failure to develop pretend play competence would leave the commitment architecture untested in low-stakes mode before deployment in high-stakes real operation." testable_prediction: "Agents whose developmental curriculum includes pretend play will show significantly fewer commitment architecture failures (false commits, commit-without-evaluation, frame confusion between real and hypothetical) in adult real-consequence operation, compared to agents whose curriculum skips pretend play but is otherwise identical. The pretend-play-skip group should show the specific failure signature of treating real consequences as if they were synthetic (frame tag confusion) or failing to commit when commitment is warranted (undertrained commitment gate)." suggested_experiment_type: developmental_ablation_discriminative_pair - id: MECH-199 title: "The caregiver's role during childhood is specifically play-frame maintenance (ARC-049), transitioning from unilateral frame-setting through co-participation to scaffolded withdrawal as play types progress." claim_type: mechanism_hypothesis subject: play.caregiver_frame_maintenance polarity: asserts status: candidate depends_on: - INV-043 - ARC-049 - ARC-046 - MECH-197 location: docs/architecture/developmental_curriculum.md notes: "The caregiver role transitions across play types: damage protection during infancy (ARC-046) -> unilateral frame-setting during sensorimotor play -> frame monitoring and boundary enforcement during constructive/pretend play -> co-participation in games with rules -> scaffolded withdrawal during cooperative play -> peer-level mutual frame maintenance in adulthood. The caregiver IS the bilateral frame-maintainer that ARC-049 requires during the developmental period when the agent cannot yet self-monitor frame integrity. This maps INV-043 (caregiver requirement) onto specific play-type phases and explains what the caregiver is concretely doing during childhood beyond protection." - id: INV-061 title: "Derealization, delusion, PTSD flashbacks, anxiety, mania, impulsivity, and commitment paralysis share a common architectural substrate: undertrained real/synthetic frame distinction from insufficient pretend play." claim_type: invariant subject: psychiatric.frame_confusion_unified_etiology polarity: asserts status: candidate depends_on: - MECH-094 - MECH-198 - MECH-200 - MECH-201 - MECH-202 location: docs/architecture/psychiatric_failure_modes.md notes: "The REE architecture distinguishes real from synthetic/hypothetical frames via the hypothesis tag (MECH-094, agent-internal) and play frame tag (ARC-049, bilateral). Pretend play (MECH-198) is where these co-operate for the first time. When pretend play is absent or insufficient, the frame-distinction mechanism is undertrained. The resulting adult failure modes are not independent disorders but different failure signatures of the same mechanism, varying by (a) direction of confusion: real->synthetic (derealization cluster, MECH-200) vs synthetic->real (delusion cluster, MECH-201), and (b) affected subsystem: frame tag, commitment gate (MECH-202), or threat evaluation. This provides a unified developmental etiology for conditions that are clinically distinct but architecturally related." - id: MECH-200 title: "Frame confusion (real->synthetic direction): the agent treats real consequences as synthetic, producing derealization, manic commitment, and dissociative states." claim_type: mechanism_hypothesis subject: psychiatric.frame_confusion_derealization polarity: asserts status: candidate depends_on: - MECH-094 - MECH-198 - ARC-049 location: docs/architecture/psychiatric_failure_modes.md notes: "The real/synthetic distinction was never properly calibrated during pretend play (MECH-198). The agent defaulted to tagging ambiguous frames as synthetic (the safer error during play, where real consequences are absent). In adulthood this default persists: when frame signals are ambiguous, real harm is discounted as if play-harm. Signatures: (1) derealization -- real harm experienced as unreal, z_harm_s fires but is discounted; (2) manic episodes -- play frame leaks into real operation, commitment gate fires without harm-gate check, elevated commitment with reduced consequence-sensitivity; (3) dissociative states -- frame tag intermittently flips to synthetic under stress. Developmental evidence: reduced pretend play predicts later psychotic symptoms (Cannon 2002)." - id: MECH-201 title: "Frame confusion (synthetic->real direction): the agent treats simulation, replay, and imagination as real, producing delusion formation, PTSD flashbacks, and anxiety." claim_type: mechanism_hypothesis subject: psychiatric.frame_confusion_delusion polarity: asserts status: candidate depends_on: - MECH-094 - MECH-198 - MECH-121 - ARC-049 location: docs/architecture/psychiatric_failure_modes.md notes: "During pretend play, the hypothesis tag (MECH-094) must be active (stick is not a sword) while the play tag (ARC-049) is also active (fight is not dangerous). If pretend play was insufficient, the hypothesis tag was never co-exercised with the play tag. In adulthood, internally generated representations lack the hypothesis tag and are processed through the real-consequence pipeline. Signatures: (1) delusions -- internal representations experienced as external reality, E3 treats simulated trajectories as real percepts; (2) PTSD flashbacks -- SWR replay (MECH-121) lacks hypothesis tag, experienced as real re-occurrence; (3) anxiety disorders -- pre-commit simulation triggers z_harm_a as if real; (4) paranoid ideation -- simulated other-agent threat (MECH-127 counterfactual) processed without hypothesis tag." - id: MECH-202 title: "Commitment gate developmental failure: undertrained gate from insufficient pretend play produces either impulsivity (gate fires without evaluation) or commitment paralysis (gate never fires)." claim_type: mechanism_hypothesis subject: psychiatric.commitment_gate_failure polarity: asserts status: candidate depends_on: - MECH-198 - MECH-197 - INV-055 location: docs/architecture/psychiatric_failure_modes.md notes: "The commitment gate was never exercised in synthetic mode during pretend play. Two failure directions depending on how the untrained gate resolves: (A) insufficient inhibition -- gate fires without completing E3 evaluation. Maps to impulsivity (commits without evaluation), mania commitment component (gate threshold too low + MECH-200 frame confusion compound), impulse control disorders (gate does not wait for E2 rollout). (B) excessive inhibition -- gate never fires, threshold stuck at infant default (INV-055 very high commit threshold never calibrated down). Maps to OCD checking/rumination (E3 loops without committing), catatonia (gate frozen at max threshold), severe indecision (commitment signal never reaches threshold). Pretend play calibrates the gate: low-stakes commitment practice adjusts the threshold between these extremes." - id: MECH-203 title: serotonergic_replay_salience_tagging claim_type: mechanism_hypothesis subject: Tonic serotonergic (5-HT) level during waking tags benefit-relevant experiences for preferential replay during subsequent SWS, complementing harm-salience tagging by the residue field polarity: positive status: candidate implementation_phase: v3 depends_on: - MECH-186 - MECH-121 - MECH-165 - SD-017 notes: > The residue field already marks harm-dense experiences for replay prioritization (MECH-099, supported by O'Neill 2010 and Huelin Gorriz 2023). But benefit-salient experiences -- where benefit_exposure was high but z_goal was weak (missed opportunities) -- should also be prioritized for consolidation. The tonic 5-HT benefit gradient (MECH-186) provides the tagging signal: experiences occurring during high tonic 5-HT (active benefit gradient) are tagged for replay, while experiences during low tonic 5-HT (depressive attractor, INV-053) are under-tagged, contributing to the consolidation deficit in depression. This connects the psychiatric failure mode (MECH-186 deficit -> reduced benefit gradient) to the sleep failure mode (under-consolidated benefit-relevant experiences -> impoverished goal representations after sleep). Predicts: depressed agents should show reduced benefit-replay density in SWS relative to harm-replay density. - id: MECH-204 title: serotonergic_rem_gate_zero_point claim_type: mechanism_hypothesis subject: Serotonergic withdrawal during REM actively defines the precision recalibration zero-point by removing the benefit gradient, not merely permitting recalibration polarity: positive status: candidate implementation_phase: v3 depends_on: - MECH-123 - MECH-186 - MECH-178 - INV-045 notes: > MECH-123 (REM precision recalibration) is currently modeled as a permissive process -- REM allows precision priors to reset. But the mechanism by which the recalibration target is established is unspecified. This claim proposes that tonic 5-HT withdrawal during REM (dorsal raphe quiescence) actively sets the zero-point: by removing the benefit gradient that biases waking precision, the system can recalibrate against a neutral reference state. This makes REM recalibration not just "allowed by" 5-HT withdrawal but "calibrated by" it. Connects to INV-045 (phase ordering): SWS must precede REM because consolidation requires the benefit gradient to be active (tagging), while recalibration requires it to be absent (zero-point). The serotonergic state transition (high tonic -> withdrawal) is part of what implements the SWS-to-REM computational boundary. Also connects to MECH-178: if noradrenergic tone suppresses REM (blocking 5-HT withdrawal), then both the replay-salience tagging (MECH-203) and the recalibration zero-point (this claim) are disrupted -- a double hit explaining why noradrenergic-mediated REM suppression produces rigidity rather than just reduced recalibration. Predicts: pharmacological 5-HT agonists during REM should impair precision recalibration quality specifically (not SWS consolidation). # Registered 2026-04-06 — surprise-gated generative replay # Source: conversation on hippocampal replay scheduling priority - id: MECH-205 title: "Hippocampal replay is surprise-gated and generative: episodes with high unexplained prediction error are prioritized, replayed with counterfactual variations via E2, and deprioritized only when forward models can account for the surprise." claim_type: mechanism_hypothesis subject: hippocampal.surprise_gated_generative_replay polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-092 - MECH-094 - MECH-121 - INV-049 - SD-003 - ARC-031 location: docs/architecture/sleep/offline_phases.md#mech-205 source: - conversation_2026-04-06_surprise_gated_replay evidence_quality_note: > Tier 1 implementation (2026-04-07): PE EMA tracking + VALENCE_SURPRISE population on residue field + PE-weighted surprise_weight in drive_state for replay start selection. Config: surprise_gated_replay=True. Write path fix (2026-04-09): EXQ-258 FAIL root-caused to two bugs -- (1) experiment script checked nonexistent _rbf_layer attr (should be rbf_field), (2) pe_ema_alpha=0.1 tracked PE so fast surprise stayed near zero. Fix: pe_ema_alpha configurable (default 0.02), pe_surprise_threshold gate (default 0.001), _surprise_write_count diagnostic. Smoke test confirms 64/100 steps produce writes with default config. EXQ-258a queued. Awaiting V3 substrate evidence (hold_pending_v3_substrate). notes: > MECH-092 establishes that quiescent E3 heartbeat cycles trigger hippocampal SWR-equivalent replay, and MECH-121 specifies the full NREM consolidation pipeline. Neither specifies what determines WHICH episodes are replayed or HOW MUCH replay each receives. This claim fills that gap: the hippocampus maintains a surprise buffer -- a priority queue of cached episodes ranked by unexplained prediction error magnitude (E1 sensory, E2 motor-sensory, or E3 harm/goal error). During offline phases, the highest-surprise episode is popped from the buffer, and the hippocampus generates counterfactual VARIATIONS around it via E2 rollouts (not rote re-presentation of the raw experience). The forward models (E1/E2) train on these replayed variations. Replay of a given episode tapers when prediction error on it and its counterfactual neighbourhood drops below threshold -- a natural convergence-based termination criterion. This makes replay allocation adaptive: more surprising events receive more replay cycles. Three functional consequences: (1) Efficient offline compute -- replay focuses on what the model gets wrong, consistent with the SWR literature showing replay bias toward novel and reward-adjacent sequences (Olafsdottir et al. 2018). (2) Generative, not rote -- the variation generation is E2 running counterfactual rollouts offline, which directly trains SD-003 attribution capacity without acting-on-incomplete-models risk (INV-049). (3) Natural termination -- events that resist explanation get more cycles, which is why they consolidate more strongly (and clinically, why traumatic events that cannot be integrated keep replaying intrusively -- the termination criterion is never met, MECH-094 tag may degrade under sustained high-affect replay). When the surprising event involved harm, replay must also continue until the residue field (viability map) updates to reflect the newly learned danger. This means the replay target is not purely forward-model prediction error but also includes residue field update completeness -- the driver is surprise, but the scope includes both model correction and map consolidation. MECH-094 hypothesis tag is critical: every generated variation must carry the simulation flag. Tag loss during generative replay = confabulation (replayed variations remembered as real). Tag loss on high-affect replays = PTSD intrusion mechanism (the replay loses its "this is simulation" marker). Architectural implication: the surprise buffer is a distinct structure from the viability map -- it is a replay SCHEDULING structure, not a spatial map. It lives in the hippocampal module and is populated during waking (high prediction error events cached with their error magnitude) and consumed during offline phases. Predicts: (1) replay frequency of an episode correlates with its initial prediction error magnitude; (2) replay of an episode decreases as the forward models improve on it; (3) blocking replay of high-surprise episodes (e.g. targeted memory reactivation of low-surprise items) should impair model correction for the blocked episodes specifically. Contrastive causal structure of replay (2026-04-07 refinement): the generative variations are not sampled arbitrarily -- they are CONTRASTIVELY paired with the anchor episode. The anchor is the surprising event (e.g. "I was hurt here" or "I reached the goal here unexpectedly"). The variations are nearby episodes where the surprising outcome DID NOT occur. The learning signal is the feature contrast: what distinguishes (context, action) pairs that produced the surprise from those that did not? This is offline causal inference, not just prediction-error minimisation. The system is asking "which antecedent features are causally upstream of this outcome?" rather than just "can I predict this outcome better?" This sharpens the termination criterion: replay does not stop when prediction error drops but when the causally relevant features have been extracted -- i.e. when the model can reliably classify surprising vs non-surprising variations by their input features. The contrastive structure is directionally asymmetric by valence: harm surprises generate variations where harm does not occur ("what would I have had to do differently to avoid this?") -- result writes to the avoidance/residue map. Goal surprises generate variations where the goal is not reached ("what were the lucky features I should exploit?") -- result writes to the approach map. Same contrastive machinery, different downstream targets. This connects directly to ARC-031 (approach-avoidance symmetry) and implements SD-003 offline: the hippocampus runs the counterfactual population E2(z_t, a_cf) across many variations simultaneously, yielding a richer causal signature than the single online comparison. The anchor episode is the observed outcome; the variation population is the counterfactual reference set; the extracted features are the offline causal attribution. Refinement C (2026-04-07, Gupta 2010 + Stoianov 2022): the generative variations are produced by the hippocampus's own hierarchical sequential model (items -> sequences -> maps), not by arbitrary E2 rollouts. The hippocampus recombines experienced trajectory segments (forward + reverse) at the sequence level to compose structurally plausible counterfactuals that were never directly traversed. E2 provides the transition dynamics; the hippocampal sequence model provides the structural constraints that keep variations coherent with the learned map. The "what if" variations are hippocampally-constrained recombinations, not unconstrained rollouts. Refinement D (2026-04-07, Mattar & Daw 2018): replay priority is not raw PE magnitude but PE x viability-relevance -- the product of gain (how much does updating this state change the agent's policy?) and need (how likely is this state to be relevant to the current viability landscape?). A highly surprising event in a region the agent will never revisit has zero need and is not prioritised despite high gain. The surprise buffer is therefore a priority queue on PE x relevance, not PE alone. This is a tighter and more normatively justified criterion than raw PE magnitude. # Registered 2026-04-07 -- CA1 comparator as surprise buffer write mechanism # Source: Duncan 2011 + Bein 2020 lit-pull synthesis for MECH-205 - id: MECH-206 title: "CA1 acts as a PE-proportional comparator that writes episodes to the surprise buffer: mismatch between CA3 prediction and entorhinal actual input produces a graded signal that tags the episode for offline replay and suppresses retrieval of the now-erroneous prediction." claim_type: mechanism_hypothesis subject: hippocampal.ca1_surprise_buffer_write_mechanism polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-205 - MECH-092 - SD-003 location: docs/architecture/sleep/offline_phases.md#mech-206 source: - evidence/literature/targeted_review_mech_205/entries/2026-04-07_mech_205_ca1_mismatch_detector_duncan2011/record.json - evidence/literature/targeted_review_mech_205/entries/2026-04-07_mech_205_mnemonic_pe_hippocampal_states_bein2020/record.json notes: > MECH-205 posits a surprise buffer (priority queue of high-PE episodes) but does not specify the biological mechanism by which episodes are tagged and written to that buffer. This claim fills that gap: CA1 is the write mechanism. CA1 receives two simultaneous inputs: CA3's pattern-completed prediction of the current context (what was expected), and entorhinal cortex's representation of the actual incoming experience (what happened). CA1 activity reflects the degree of discrepancy between these two channels -- automatically, graded, and proportionally (Duncan et al. 2011: CA1 tracks total number of changes linearly regardless of task relevance). When mismatch is high, a bidirectional connectivity shift occurs (Bein et al. 2020): CA1-entorhinal connectivity increases (encode the surprising input) while CA1-CA3 connectivity decreases (suppress retrieval of the now-erroneous prediction). This implements the surprise buffer write in two steps: (1) the episode is tagged with its PE magnitude via CA1 mismatch signal; (2) retrieval of the old prediction is suppressed to prevent the erroneous model from continuing to drive behaviour. The graded, linear sensitivity is critical: the surprise buffer needs a continuous PE magnitude readout to rank episodes for replay priority. A binary match/mismatch flag would not support the priority ordering MECH-205 requires. CA1's graded mismatch signal is exactly the right quantity. During SWR-driven offline replay, the same CA1 comparator logic operates in the generative direction: CA3 pattern-completes the anchor episode; hippocampally-generated variations arrive via entorhinal-equivalent input; CA1 computes the structural difference between anchor and each variation. This is the contrastive comparison signal that drives causal feature extraction (MECH-205 contrastive structure). The CA1 mismatch signal during replay identifies which features distinguish anchor from variation, providing the learning signal for forward model correction. REE architectural implication: the HippocampalModule needs a CA1-equivalent comparator submodule receiving (i) CA3-like pattern-completed anchor representations and (ii) entorhinal-like actual/variation representations, with output being a graded PE signal that writes to the surprise buffer priority queue. Predicts: (1) disrupting CA1 activity specifically (not CA3 or DG) should impair surprise buffer population while leaving forward replay intact; (2) the CA1 mismatch signal during replay should correlate with which features the agent subsequently treats as causally relevant to the surprising outcome. # Registered 2026-04-07 -- cholinergic write-gate on surprise buffer # Source: Sinclair 2021 lit-pull synthesis for MECH-205 - id: MECH-207 title: "Acetylcholine acts as a permissive write-gate on the surprise buffer: prediction errors trigger hippocampal representation destabilisation and offline memory updating only when basal forebrain cholinergic activation co-occurs." claim_type: mechanism_hypothesis subject: hippocampal.cholinergic_surprise_buffer_gate polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-205 - MECH-206 - MECH-178 location: docs/architecture/sleep/offline_phases.md#mech-207 source: - evidence/literature/targeted_review_mech_205/entries/2026-04-07_mech_205_prediction_error_memory_sinclair2021/record.json notes: > MECH-205 treats the surprise buffer as populated by any high-PE episode. Sinclair et al. 2021 (PNAS) show that prediction errors trigger hippocampal destabilisation and memory updating ONLY when basal forebrain cholinergic activation co-occurs. Without ACh, the PE signal does not make memories labile for updating, even if the PE is large. This implies a neuromodulatory write-gate on the surprise buffer: ACh is the permissive signal that determines which PE-tagged episodes actually get written to the buffer. PE alone is necessary but not sufficient -- the agent must be in an ACh-permissive state (aroused, attending) for the episode to be cached. This explains: (1) inattentional learning failure -- you can be objectively surprised (high PE) and fail to update if ACh is low (drowsy, inattentive); (2) novelty enhances learning -- novel contexts elevate ACh (basal forebrain is strongly novelty-responsive), opening the write-gate for PE episodes encountered in that context; (3) the 24-hour delay in updating (Sinclair 2021) is consistent with offline sleep-phase consolidation of ACh-gated cached episodes (MECH-205 / MECH-121). Clinical implication: ACh depletion (as in Alzheimer's disease, where basal forebrain is an early casualty) closes the surprise buffer write-gate even when the CA1 comparator (MECH-206) generates normal PE signals. The forward models accumulate PE without updating -- the patient registers that something is wrong but cannot integrate the correction. A plausible mechanism for confabulation and inertia-of-false-belief in early AD. Two-gate architecture: NE controls when offline replay occurs (via REM gating, MECH-178); ACh controls which episodes are eligible for offline updating (surprise buffer gate). They are independent axes: REM suppression (high NE) blocks the offline phase; low ACh at encoding blocks which episodes enter the buffer in the first place. Both gates must be functional for a PE episode to reach offline correction. Implementation_phase v4: requires neuromodulatory layer in hippocampal module. Predicts: (1) ACh blockade at time of PE prevents offline updating of that episode even if replay occurs; (2) ACh agonists at PE time enhance offline updating; (3) in AD, PE-to-updating delay is disproportionately impaired relative to PE detection. # Registered 2026-04-07 -- valence-asymmetric replay and anxiety mechanism # Source: McFadyen 2023 lit-pull synthesis for MECH-205 - id: MECH-208 title: "Valence-asymmetric replay causally drives approach/avoidance bias: harm-path replay is weighted relative to reward-path replay by the agent's harm-salience parameter, and chronically elevated harm-weighting produces avoidance bias in safe contexts." claim_type: mechanism_hypothesis subject: hippocampal.valence_asymmetric_replay_approach_avoidance polarity: asserts status: candidate implementation_phase: v3 depends_on: - MECH-205 - ARC-031 - MECH-186 - MECH-124 location: docs/architecture/sleep/offline_phases.md#mech-208 source: - evidence/literature/targeted_review_mech_205/entries/2026-04-07_mech_205_valence_asymmetric_replay_mcfadyen2023/record.json notes: > MECH-205 claims the contrastive replay structure is valence-asymmetric (harm surprises write to avoidance map, goal surprises write to approach map). McFadyen et al. 2023 (Nature Neuroscience) show this asymmetry is not merely structural but causally drives downstream decisions, and is modulated by individual differences in harm-weighting. Replay of punishment paths was boosted relative to reward paths before avoidance decisions, and attenuated before approach decisions. Trial-by-trial bias toward punishment-path replay predicted irrational approach decisions -- excess threat evidence tipped the decision toward avoidance even in safe contexts. The effect was stronger in high trait-anxiety individuals. In REE terms: the harm-salience parameter (z_beta / residue field weighting) modulates the valence asymmetry in replay. Elevated harm-salience causes punishment-path replay to dominate, systematically biasing the contrastive comparison toward harm-relevant features. The agent extracts harm-predictive features from a larger fraction of its replay budget than is warranted by actual harm probability -- a miscalibration that generalises to safe contexts, producing avoidance of situations that do not warrant it. This is the replay-mediated anxiety mechanism: not that anxious agents perceive more harm in the moment, but that they replay harm-relevant paths more heavily during offline consolidation, installing a harm-weighted cognitive map that biases future decisions. The mechanism is self-reinforcing: elevated harm-weighting -> more harm-path replay -> more harm-predictive feature extraction -> more avoidance -> fewer approach experiences -> less corrective evidence -> maintained harm-weighting. This is the consolidation- mediated option-space contraction of MECH-124, specified here as operating via the valence asymmetry in replay content. Connects to MECH-186 (psychiatric failure mode cluster). The beta-blocker-before-sleep prediction follows directly: propranolol reduces NE-mediated harm salience during consolidation, rebalancing the replay ratio toward reward paths, attenuating avoidance bias. Consistent with the literature on propranolol and fear memory reconsolidation. Predicts: (1) disrupting harm-path replay specifically should attenuate avoidance behaviour in anxious agents without affecting approach in safe contexts; (2) the ratio of harm-path to reward-path replay during offline phases predicts trait anxiety scores; (3) interventions rebalancing the replay ratio (exposure therapy, beta-blockers pre-sleep) reduce avoidance bias by correcting the harm-weighted cognitive map installed by asymmetric replay. # Registered 2026-04-07 -- dream phenomenology as window onto replay type # Source: Daniel Golden first-person observation; three-type dream taxonomy - id: INV-062 title: "Dream phenomenology is a first-person window onto active offline sleep process: four structurally distinct dream types map onto four computationally distinct sleep functions -- harm-contrastive replay, goal-contrastive replay, NREM semantic consolidation, and REM world-model free-running." claim_type: invariant subject: dream_phenomenology.replay_type_signature polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-205 - MECH-208 - MECH-209 - MECH-210 - INV-049 - MECH-121 - MECH-123 location: docs/architecture/sleep/offline_phases.md#inv-062 source: - docs/thoughts/2026-04-07_dream_phenomenology_replay_types.md - evidence/planning/thought_intake_2026-04-07_dream_phenomenology_replay_types.md notes: > Four structurally distinct dream types observed consistently by DG (first-person phenomenological data; noted as definite types but not claimed to be exhaustive): Type 1 (stress/escape frustration): trying to get away from something; everything that could go wrong does go wrong. No resolution. Type 2 (joy/goal frustration): trying to reach a goal; everything that could go wrong does go wrong. No resolution. Affect is different from Type 1 (approach vs avoidance) but the systematic-failure structure is identical. Type 3 (procedural superposition): one thing experienced from all angles, all scales, and all times simultaneously -- opening a door that is all kinds of doors at once; a cherry blossom tree blooming experienced from all perspectives, scales, and moments at once. No agent, no goal, phenomenologically peaceful. Type 4 (detailed story / close to reality): coherent narrative with full story arc; close to real-world plausibility; novel situations not directly experienced; super interesting. An agent present in a coherent world where things follow causally from each other. Four-type mapping to computational processes: Types 1 and 2: harm-contrastive and goal-contrastive replay (MECH-205) not converging. The systematic-failure structure is the generative model exhaustively sampling the failure space -- variations that all hit the failure condition. Frustration IS the convergence failure signal; non-resolution is a phenomenological report that the replay loop is still running. Mirror images by valence (MECH-208 / ARC-031). Type 3: NREM SWR semantic consolidation (MECH-121 / MECH-209). Rapid SWR sequence collapses to phenomenological simultaneity. Self-dissolution = episodic context stripped during abstraction. Multi-scale = hierarchical model across levels. Peaceful because there is no convergence criterion -- building a schema, not solving a problem. Type 4: REM E1 world model free-running (MECH-210 / MECH-123). Sensory input disconnected, commit gate suppressed; E1 LSTM runs forward unconstrained on its learned prior. Sequential LSTM dynamics produce narrative arc. "Close to reality" = in- distribution samples from a well-calibrated world model. "Super interesting" = E1 surfacing novel-but-coherent patterns from its prior that the agent hasn't consciously accessed -- world model hypothesis testing. Agent present in a coherent world because E1 is a sequential agent-world model, not a categorical schema extractor. Cross-type comparison: Types 1/2 and 4 all have an agent; Type 3 does not. Types 1/2 have no narrative resolution (convergence criterion unmet); Type 4 has narrative arc (no convergence criterion -- free generation). Type 3 has no agent and no narrative (schema building). The presence/absence of agent and resolution criterion cleanly discriminates all four. Clinical implications: (1) Recurring non-resolving Type 1 = convergence failure on harm-contrastive replay; PTSD marker; should resolve with successful therapy. (2) Absence of Type 3 = possible MECH-121 failure (episodic->semantic transfer impaired). (3) Absence of Type 4 or degraded Type 4 (incoherent, implausible stories) = E1 world model poorly calibrated or REM architecture disrupted (MECH-123 / MECH-178). (4) Anxious individuals (elevated harm-weighting, MECH-208) predicted to have predominantly Type 1 with reduced Type 3 and Type 4 -- replay budget crowded by harm-contrastive search. Epistemic note: first-person phenomenological data from one observer. Taxonomy observed, not derived. The mapping to computational processes is an interpretation requiring corroboration from other observers and sleep-stage measurement. # Registered 2026-04-07 -- Type 3 dream as semantic consolidation phenomenology # Source: Daniel Golden first-person observation; MECH-121 lit synthesis - id: MECH-209 title: "The simultaneous multi-instance, multi-scale, multi-time phenomenology of Type 3 dreams is the experiential signature of NREM SWR episodic-to-semantic transfer: rapid hippocampal sequential presentation of concept instances collapses to phenomenological simultaneity, with self-dissolution reflecting the stripping of episodic context during abstraction." claim_type: mechanism_hypothesis subject: sleep.semantic_consolidation_phenomenology polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-121 - MECH-092 - INV-062 - INV-049 location: docs/architecture/sleep/offline_phases.md#mech-209 source: - docs/thoughts/2026-04-07_dream_phenomenology_replay_types.md - evidence/planning/thought_intake_2026-04-07_dream_phenomenology_replay_types.md notes: > MECH-121 describes NREM SWR replay as the mechanism for episodic-to-semantic transfer: hippocampal SWR-equivalent replay coordinates with slow oscillations to deliver episodic content to neocortex (E1's LSTM) for schema integration. The computational process involves rapid sequential presentation of concept instances (compressed to 20-90ms per state, McFadyen 2023), allowing E1 to abstract the invariant features across instances. This claim specifies what that process looks like from the inside, given any phenomenological access during NREM. Because the hippocampal SWR sequence operates at timescales far below conscious temporal resolution, the sequence collapses to simultaneity at the phenomenological level: all instances of the concept are present at once, rather than being experienced one after another. Concrete examples from DG's experience: (1) Opening a door experienced as all kinds of doors at once -- all instances of door in memory simultaneously present. (2) A cherry blossom tree blooming experienced from all angles, all scales (petal to branch to tree to forest), and all moments of blooming at once. Two structural features characterise Type 3: (i) MULTI-DIMENSIONALITY -- all angles, all scales, all times simultaneously, not a sequence; (ii) SELF-DISSOLUTION -- no agent trying and failing, no goal, no frustration. The dreamer IS the superposition, not an observer of it or an actor within it. Self-dissolution maps onto the episodic context stripping that defines semantic consolidation: episodic memory preserves who, when, where (the self in a situation); semantic memory strips those and retains only the what (the category's invariant features). During the SWR-driven consolidation process, the episodic binding that anchors experience to a self-in-a-situation is exactly what is being removed. The phenomenological self therefore dissolves -- not because consciousness is absent, but because the self-context that normally grounds phenomenological perspective is the content being discarded. Multi-scale access maps onto the hierarchical generative model (Stoianov 2022): items -> sequences -> maps. The dreamer simultaneously accesses all levels of the hierarchy (petal = item, branch sequence = sequence, whole forest = map). This would be experienced as "all scales at once" precisely because the hierarchical model is being consolidated top-to-bottom in a single SWR burst. Multi-time maps onto the temporal abstraction required for schema formation: the schema for "cherry blossom blooming" must abstract across all the times the dreamer has observed blooming, presenting them simultaneously to E1 for invariant extraction. Therapeutic / clinical implication: if Type 3 dreams are the phenomenological signature of healthy semantic consolidation, then their absence (or displacement by Type 1/2 dreams crowding the replay budget) would indicate impaired episodic->semantic transfer. In chronic anxiety (MECH-208): elevated harm-weighting allocates excessive replay budget to harm-contrastive search (Type 1), leaving insufficient budget for semantic consolidation (Type 3). Clinically: anxious patients may report predominantly Type 1/2 dreams and rarely Type 3 -- a testable prediction using prospective dream diaries. Connects to MECH-124 (consolidation-mediated option-space contraction): if harm-path replay dominates the budget, not only does the avoidance map get over-consolidated, but the semantic abstraction that would provide generalisable escape/approach schemas gets under-consolidated. # Registered 2026-04-07 -- Type 4 dream as REM E1 world-model free-running # Source: Daniel Golden first-person observation; MECH-123 synthesis - id: MECH-210 title: "Type 4 (detailed story, close to reality, super interesting) dreams are the phenomenological signature of REM E1 world-model free-running: with sensory input disconnected and the commit gate suppressed, E1's LSTM generates unconstrained sequential rollouts from its learned prior, producing coherent narrative arcs that are novel-but-plausible and experientially interesting." claim_type: mechanism_hypothesis subject: sleep.rem_e1_world_model_free_running polarity: asserts status: candidate implementation_phase: v4 depends_on: - MECH-123 - MECH-094 - MECH-178 - INV-049 - INV-062 location: docs/architecture/sleep/offline_phases.md#mech-210 source: - docs/thoughts/2026-04-07_dream_phenomenology_replay_types.md - evidence/planning/thought_intake_2026-04-07_dream_phenomenology_replay_types.md notes: > During waking, E1's LSTM is constrained by continuous sensory input -- it predicts the next state, receives the actual state, and updates on the error. During REM, three constraints lift simultaneously: (1) sensory input is disconnected (thalamocortical gating); (2) the commit gate is fully suppressed (ARC-016 running_variance threshold effectively infinite -- no committed action can occur); (3) monoaminergic tone (noradrenaline, serotonin) withdraws, reducing precision on prior predictions and allowing the model to explore more freely. E1 is thus free to run its LSTM forward from any internally-generated starting point, following its learned transition model without correction from the world. All activity is hypothesis-tagged (MECH-094). The result is that E1 generates what is effectively a story: a sequential, causally coherent rollout of world states, with the agent as a participant whose actions have consequences within the simulation. Because E1's prior is learned from real experience, the rollouts stay within the distributional support of plausible events -- they are "close to reality." Because they are samples from the prior rather than memories, they are novel -- situations and event sequences the agent has not experienced but that E1's world model treats as possible. The "super interesting" quality is the key diagnostic marker. It signals that E1 is surfacing novel-but-coherent structure from its learned prior: causal chains, social dynamics, physical sequences that the model has implicitly learned but that have not been consciously encountered as episodes. This is MECH-123's "world model hypothesis testing": E1 explores whether its current prior produces internally consistent trajectories. When the prior is well-calibrated, the rollouts are interesting because they reveal structure that was latent but not yet explicit. When the prior is miscalibrated, the rollouts are either boring (repetitive, low-information samples) or incoherent (impossible, contradictory sequences). Narrative arc emerges naturally from E1's sequential LSTM dynamics with prediction_horizon=20: the model generates extended temporal sequences in which earlier states constrain later ones, producing setup, development, and consequence -- the structural signature of story. This is not imposed on the model; it is what a sequential world model with a long prediction horizon produces when run forward freely. Contrast with other dream types: unlike Types 1/2 (hippocampal contrastive replay, frustration from convergence failure), Type 4 has no convergence criterion -- E1 is generating, not searching. Unlike Type 3 (NREM semantic consolidation, self-dissolved superposition), Type 4 has a present agent in a sequential world -- because E1 is an agent-world model, not a categorical schema extractor. The distinction between Type 4 and Type 1/2 in terms of frustration is also mechanistically grounded: in Types 1/2, the hippocampus is iterating variations that keep failing, and the emotional content of the anchor episode (harm, blocked goal) sets the affective tone. In Type 4, E1 is free-running from an arbitrary starting point with no failure criterion -- there may be challenges in the narrative, but the dreamer can engage with them as interesting problems rather than frustrated obstacles, because there is no external convergence criterion demanding resolution. Clinical implications: (1) degraded Type 4 (incoherent stories, impossible events, implausible causality) = E1 world model poorly calibrated or REM integrity disrupted (MECH-123, MECH-178). (2) Absence of Type 4 in patients on REM-suppressing medications (e.g. many antidepressants suppress REM) is a testable prediction -- they should lose Type 4 while Types 1-3 may be less affected. (3) The subjective quality of Type 4 dreams (how interesting, how coherent, how novel) may be an accessible proxy for E1 world model quality -- richer and more interesting stories suggesting a better- calibrated world model. (4) In dementia: progressive loss of world model fidelity (MECH-178 pathway, rigidity mechanism) predicts progressive degradation of Type 4 towards repetitive, low-information, close-to-habit dreams before Type 4 disappears entirely -- a gradient that dream diaries might track ahead of other symptoms. - id: INV-063 title: minimum_entropy_intake_sleep_dependency claim_type: invariant status: candidate subject: sleep.entropy_intake description: > A minimum level of environmental entropy intake during waking is required to sustain all four offline sleep functions. Too little entropy collection produces specific, proportional sleep function failures: insufficient schema material starves NREM Type 3 consolidation; insufficient surprising episodes starve Types 1/2 contrastive replay; insufficient novel event sequences starve E1 world-model updating; insufficient high-entropy sensory input degrades E2 motor-sequence learning. This predicts a psychiatric risk gradient: environments with low entropy (straight lines, routine, little novelty — characteristic of urban and institutional settings) systematically deplete the material for sleep function, producing progressive failure modes across anxiety, depression, and cognitive rigidity dimensions. High-entropy environments (nature, fractal structure, novel sensory complexity) provide the necessary raw material for all four functions and are predicted to be protective. notes: > Motivated by Daniel Golden's observation that nature exposure (fractal visual complexity, unpredictable structure) and urban deprivation (straight lines, routine, low novelty) differentially affect psychiatric outcomes in ways not explained by current accounts. The entropy-intake mechanism links waking experience quality to sleep function completeness: the brain cannot consolidate, replay, or recalibrate what it has not encountered. Minimum entropy thresholds are likely E1-specific (world-model schema update requires novel world states) and E2-specific (motor sequence learning requires action-outcome variation). This is a novel prediction that bridges environmental psychiatry, chronobiology, and the REE sleep framework. registered: "2026-04-07" depends_on: [INV-049, MECH-121, MECH-205, MECH-209, MECH-210, MECH-211] - id: MECH-211 title: schema_consolidation_as_search_grammar claim_type: mechanism status: candidate subject: sleep.nrem.schema_vocabulary description: > Type 3 NREM semantic consolidation (SWR-driven episodic-to-semantic transfer) produces the generative vocabulary that Types 1 and 2 contrastive replay searches draw upon for variation generation. Schemas consolidated during Type 3 sleep provide categorical structure — object types, action primitives, contextual frames — that constrains and enriches the counterfactual space sampled by E2 during Types 1/2 replay. Richer Type 3 consolidation therefore enables a broader, better-structured contrastive search space in subsequent Types 1/2 episodes. Depleted Type 3 consolidation (insufficient schema material from waking, or hippocampal episodic-to-semantic transfer failure) narrows the variation space, producing stereotyped, low-diversity replay that fails to identify the causal features distinguishing safe from unsafe contexts — chronic convergence failure. notes: > This makes NREM consolidation a prerequisite for contrastive replay quality, not merely a parallel process. Type 3 is not just semantically restful; it actively arms the E2 variation generator with the vocabulary for its search. MECH-205's convergence criterion depends on variation diversity; MECH-211 identifies where that diversity comes from. Clinical prediction: patients with impaired NREM spindle activity (a marker of hippocampal-neocortical transfer) should show more stereotyped, narrow contrastive replay — consistent with the rigid, repetitive quality of trauma-related nightmares in PTSD. registered: "2026-04-07" depends_on: [MECH-205, MECH-209, INV-062] - id: MECH-212 title: e2_motor_sequence_sleep_consolidation claim_type: mechanism status: candidate subject: sleep.nrem.e2_consolidation description: > E2 (the fast motor-sensory transition model) undergoes dedicated NREM offline consolidation analogous to but distinct from E1 schema consolidation. While E1 consolidates world-model schemas (what objects and events exist), E2 consolidates action-sequence schemas (how to traverse z_self transition space efficiently): chains of action-objects that reliably navigate from initial to target z_self states. This E2 NREM replay is driven by the same SWR mechanism but carries motor-sensorium prediction error rather than world-state prediction error. Disrupted E2 consolidation predicts impaired skill acquisition and procedural memory — the well-established sleep-dependent motor learning effect (Walker 2005, Fischer 2002) is the phenomenological signature of E2 NREM replay. In REE terms: SWRs during NREM replay high-PE E2 state transitions; the surprise buffer for E2 is populated by action sequences that produced unexpectedly large motor-sensory prediction error (clumsy, novel, or skill-boundary episodes). notes: > This is a significant gap in the current sleep architecture (MECH-120 through MECH-123), which is E1-centric. E2 has no dedicated offline consolidation claim despite being a full transition model with its own prediction error signal and therefore its own surprise buffer population mechanism. Motor sequence learning is one of the most replicated sleep-dependent memory findings; its absence from the REE sleep account is an architectural omission. E2's action_objects become the hippocampal map backbone (SD-004), so E2 consolidation may also feed the E3 viability map indirectly. registered: "2026-04-07" depends_on: [MECH-205, MECH-121, INV-049] implementation_phase: v3 - id: MECH-213 title: e1_e2_joint_rem_calibration claim_type: mechanism status: candidate subject: sleep.rem.joint_calibration description: > REM phase involves joint E1-E2 calibration, not E1 recalibration alone. E1 generates world-state sequences (narrative context) while E2 simultaneously generates action sequences through those world states (the agent acting within the narrative). The two models run in a partially decoupled closed loop: E1's world evolution is shaped by E2's action proposals, and E2's action-value estimates are updated by E1's predicted world responses. The phenomenological signature of this joint operation is the Type 4 dream: a coherent story with an agent navigating it — neither pure sensory replay (E1 alone) nor pure motor sequence replay (E2 alone), but the joint product of both forward models running together without external grounding. The agentive quality of REM dreams (I am there, I am acting) directly reflects E2's active participation; pure E1 replay would produce observer-perspective narrative without first-person agency. This joint REM operation recalibrates the match between E1's world predictions and E2's action-consequence expectations, correcting systematic biases that accumulated during waking experience. notes: > This distinguishes REE's REM account from E1-centric accounts. MECH-123 (current REM recalibration claim) should be updated to note E2 participation. The clinical prediction: disrupted REM (sleep apnea, fragmentation, REM suppression by antidepressants) impairs E1-E2 joint calibration, producing systematic divergence between world-model predictions and motor-consequence expectations. This divergence may manifest as action-plan brittleness, impaired transfer of skills to novel contexts, or the 'known-what but not how' dissociation. SSRIs suppress REM; their side effect of emotional blunting or feeling 'disconnected' from action may partly reflect impaired E1-E2 joint calibration. registered: "2026-04-07" depends_on: [MECH-123, MECH-210, MECH-212, INV-062] implementation_phase: v3 - id: INV-064 title: e1_e2_e3_maturational_sequence_necessity claim_type: invariant status: candidate subject: development.maturational_sequence description: > The developmental sequence in which sensory-perceptual systems (E1) mature before motor-transition systems (E2) before frontal-evaluative systems (E3) is a computational necessity, not an anatomical accident. E3's primary inputs are z_world from E1 and action_objects from E2. If these inputs carry only poorly-differentiated, low-information representations, E3 training produces a noise-fitted harm/goal evaluator. The information quality available to E3 is strictly bounded by E1/E2 representational differentiation. Therefore productive E3 training cannot begin until E1 and E2 have reached sufficient schema differentiation. REE predicts that the biological fact of prefrontal cortex being the last cortical region to myelinate fully (completing mid-20s in humans) is architecturally required: any agent architecture with this dependency structure will exhibit the same maturation order. notes: > This unifies several otherwise unexplained developmental facts: why adolescent risk-taking persists well past sensory-motor competence (E1/E2 are largely mature but E3 training is still integrating), why early-life disruption has outsized and specific effects on executive function rather than just emotional regulation, and why intensive novel experience in childhood (high-entropy environments, play, exploratory learning) has disproportionate long-term cognitive returns relative to the same experience in adulthood. The maturational sequence is not merely a constraint to work around -- it is the correct curriculum order for building a competent E3. registered: "2026-04-07" depends_on: [INV-055, INV-056, MECH-212, MECH-213] - id: MECH-214 title: goal_referent_e1_representability claim_type: mechanism status: candidate subject: development.goal_coherence implementation_phase: v4 description: > The content of any goal -- the state or object being aimed at -- must be representable in E1's latent schema space for z_goal to carry semantic information. A goal aimed at an E1-unrepresented state is, functionally, a random vector in E3's operating space: it cannot generate coherent viability terrain, cannot guide hippocampal trajectory proposals, and cannot produce meaningful harm/benefit evaluation. Goal coherence therefore co-develops with E1 schema richness: primitive goals (warmth, satiation, physical contact) emerge earliest because those states are E1-representable from earliest sensory experience; higher-order social, aesthetic, ethical, and self-transcendent goals emerge only after E1 has differentiated the relevant schema vocabulary. Goals are E1 functions in this precise sense: their referent content is constituted by E1 representations, not independently specified. The frontal goal-evaluation system can only evaluate what the world-model can represent. notes: > Clinical predictions: (1) Early-life disruption of E1 schema formation (sensory deprivation, early trauma, neurodevelopmental conditions affecting sensory processing) does not merely cause emotional injury -- it constrains the semantic space from which coherent goals can ever be drawn, with downstream effects on E3 training quality that persist into adulthood. (2) Interventions that expand E1 schema richness (novel high-entropy environments, therapy that develops experiential vocabulary, practices that develop fine perceptual discrimination) can widen the space of coherent goals available to E3, not just regulate existing ones. (3) Goal incoherence -- wanting without clear object, motivation without direction -- may reflect under-differentiated E1 schemas rather than E3/reward pathology. This reframes certain anhedonic and motivationally flat presentations as E1 poverty rather than purely dopaminergic disorder. V3 WIRING AUDIT (2026-04-07): z_goal lives purely in z_world space (GoalState seeds from z_world_current). V3 grid world conflates location with reward, so z_world-only goals are adequate. V4 needs z_self-domain goal representation (DR-11) and an environment where proxy and hedonic content dissociate (DR-14) to surface this failure mode. See also addiction mapping: repeated pursuit of z_goal proxy without hedonic grounding = wanting fires on E1-unrepresented satisfaction state. registered: "2026-04-07" depends_on: [INV-064, MECH-205, MECH-121] - id: MECH-215 title: self_model_prerequisite_for_agentive_prediction claim_type: mechanism status: candidate subject: development.self_model_agency implementation_phase: v4 description: > Coherent goal-achievement prediction requires a well-differentiated self-model as the subject of that prediction. In REE terms this decomposes into two prerequisites that must precede E3 viability planning: (1) E1 self-schema differentiation -- the agent must have a stable, accurate z_self representation (body schema, proprioceptive state, affective baseline, capacity envelope) before it knows what kind of entity is doing the pursuing; (2) E2 self-transition differentiation -- the agent must have learned how its actions transform its own z_self state before it can predict whether it is capable of navigating a given path. Without (1), E3's viability trajectories have no coherent subject: they are plans for an undefined agent. Without (2), E3 cannot estimate whether the agent's affordances are sufficient to traverse the proposed path. Goal-achievement prediction is therefore prediction of form: can *this* agent, with *this* capacity model, reach *that* state? The "I" that makes a goal a first-person goal -- not just a desired world-state but a personally pursuable one -- is constituted by E1's self-schemas and E2's self-transition model, not by E3 itself. notes: > Clinical predictions: (1) Depersonalisation and derealization (disrupted E1 self-schema) should produce goal-directedness failure that looks like motivational disorder but is actually a subject-dissolution problem -- the "I" required to be the agent of pursuit is not coherently represented. (2) Conditions that degrade the body schema or interoceptive accuracy (chronic pain, fatigue, dissociative disorders) impair E2 self-transition accuracy, producing inflated or deflated capacity estimates that make viability predictions systematically wrong -- the person either overestimates what they can do (mania-like) or catastrophically underestimates (avoidance, agoraphobia, learned helplessness). (3) Somatic/body-based and interoception-focused therapies work at the correct level: they rebuild E1 self-schemas and E2 self-transition accuracy, restoring the preconditions for coherent agency rather than directly targeting E3 goal/harm evaluation. (4) Infant motor development (proprioceptive and body schema formation before goal-directed reaching) follows this sequence exactly. "Know thyself" is not ethical wisdom but an architectural prerequisite. V3 WIRING AUDIT (2026-04-07): E3.score_trajectory() evaluates entirely in z_world space -- z_self is invisible to E3 (DR-10). E3 trusts E2 rollout unconditionally -- no E2 PE -> E3 confidence modulation (DR-12). z_self encoder has no temporal depth -- body_obs -> MLP -> EMA, no recurrence (DR-13). All three gaps are correctly scoped for V3's simple 4-action grid world but become critical when the agent has complex body dynamics or the environment requires capacity-aware planning. registered: "2026-04-07" depends_on: [INV-064, MECH-214, MECH-212] - id: SD-018 title: encoder.resource_proximity_supervision claim_type: substrate_decision status: implemented subject: encoder.resource_proximity description: > z_world encoder requires auxiliary resource proximity regression loss to encode benefit-relevant features. Without this, z_world is trained only on E1 world-model prediction loss, which is invariant to resource saliency (the resource does not change the sensory scene unless contacted). EXQ-085m proved benefit_eval_r2=-0.004: z_world is orthogonal to resource proximity. A Sigmoid regression head on z_world predicting max(resource_field_view) in [0,1] backprops through the encoder, forcing z_world to represent resource proximity. This is the benefit-side analog of SD-009, which solved the same problem for harm-event discrimination. Without SD-018, the entire benefit/goal scoring pathway (benefit_eval_head, goal_proximity, z_goal seeding, drive modulation, dual systems) operates on noise. SD-018 is the critical unblocking substrate decision for V3 goal-directed behavior. notes: > Implementation: SplitEncoder.resource_proximity_head = nn.Sequential( nn.Linear(world_dim, 1), nn.Sigmoid()). Config: use_resource_proximity_head (bool, default False), resource_proximity_weight (float, default 0.5). Agent method: compute_resource_proximity_loss(target, latent_state) -> MSE loss. Target: max(resource_field_view) from CausalGridWorldV2 (available when use_proxy_fields=True). Gradient flows through z_world encoder. Backward-compatible: disabled by default. Identified during V3 wiring audit 2026-04-07: the z_world-doesn't-encode-resources gap explains all goal-directed behavior failures (EXQ-085m, EXQ-225, EXQ-237a, EXQ-238, and the entire EXQ-085h-085o SD-015 series). registered: "2026-04-07" depends_on: [SD-009, SD-005] - id: SD-019 title: affective_harm_nonredundancy_constraint claim_type: design_decision status: candidate implementation_phase: v3 subject: harm_stream.affective_nonredundancy polarity: asserts description: > The affective harm stream (`z_harm_a`) must encode temporally integrated threat load, not a delayed, smoothed, or monotone-transformed copy of the sensory-discriminative harm stream (`z_harm_s`). `z_harm_s` tracks immediate, action-conditional hazard proximity/intensity and supports forward prediction (`E2_harm_s`). `z_harm_a` serves a different function: it encodes accumulated threat burden with motivational relevance for urgency and commitment modulation. A valid `z_harm_a` implementation must therefore admit dissociations where immediate sensory harm is low but affective load remains elevated, and where brief sensory spikes do not saturate affective load. notes: > This claim sharpens SD-011. SD-011 establishes the dual-stream split in principle: sensory-discriminative vs affective-motivational. SD-019 adds a stricter substrate constraint: the affective stream is invalid if it is recoverable as a simple lagged or monotone mapping from the sensory stream. Biological grounding: the C-fiber / paleospinothalamic pathway is slower and more integrative than the A-delta pathway, but the critical computational property is not "delay" alone; it is nonredundant temporal integration with persistence and slower recovery. Motivated by EXQ-241 and EXQ-241a: both show that `z_harm_a` can track accumulated harm while remaining too close to `z_harm_s` functionally (D3 reversal persists in EXQ-241a despite harm-history input). The design requirement is therefore representational nonredundancy, not merely temporal lag. registered: "2026-04-08" depends_on: [SD-011, SD-010, ARC-016] location: docs/architecture/sensory_stream_tags.md#sd-019 - id: MECH-219 title: affective_harm_hysteretic_integration claim_type: mechanism_hypothesis status: candidate implementation_phase: v3 subject: harm_stream.hysteretic_affective_load polarity: asserts description: > `z_harm_a` is generated from a temporally extended harm-history signal by a hysteretic integration rule: affective load rises under sustained or repeated harm exposure and decays more slowly once exposure stops. The resulting signal is not reconstructible from instantaneous sensory harm alone and therefore provides E3 with a genuine motivational urgency channel rather than a second copy of proximity. The first operational form is a leaky integrator over recent harm exposure with asymmetric onset and recovery parameters. notes: > This is the first concrete mechanism proposal satisfying SD-019. A pure delay (`z_harm_a(t) = z_harm_s(t-k)`) is insufficient because it preserves recoverability from the sensory stream and is unlikely to produce meaningful motivational dissociation. The proposed mechanism instead uses hysteretic integration: repeated or sustained harm drives the load upward, while recovery is slower and path-dependent. This matches the intended role of `z_harm_a` in ARC-016-style urgency and commitment modulation. The mechanism is compatible with the harm-history input added on 2026-04-08, but makes the stronger claim that asymmetric rise/decay dynamics are necessary. Testable predictions: (1) after harm offset, `z_harm_s` falls rapidly while `z_harm_a` remains elevated; (2) repeated brief harms summate in `z_harm_a` more than single isolated harms; (3) two trajectories with matched current `z_harm_s` but different recent harm histories produce different `z_harm_a` norms and different urgency effects in E3. registered: "2026-04-08" depends_on: [SD-019, SD-011, ARC-016] location: docs/architecture/sensory_stream_tags.md#mech-219 - id: Q-036 title: "What additional variables, beyond temporal integration, are required for affective harm to become a genuinely distinct load state: persistence, recovery failure, uncontrollability, inescapability, or prediction error?" claim_type: question subject: harm_stream.affective_load_drivers polarity: open_question status: open confidence: 0.0 depends_on: - SD-019 - MECH-219 location: docs/architecture/sensory_stream_tags.md#q-036 short: > Is temporal hysteresis sufficient for `z_harm_a`, or must affective load also depend on persistence, recovery failure, uncontrollability, inescapability, or surprise? description: > Temporal integration is the most conservative mechanism for differentiating affective harm from sensory harm, but it may not be sufficient. Human affective pain and threat load appear to depend not only on duration and intensity but also on whether harm is escapable, controllable, surprising, or recurrent despite attempted avoidance. The answer determines whether MECH-219 can remain a simple hysteretic integrator or whether `z_harm_a` requires additional weighting terms. REE implication: if temporal hysteresis alone fails to produce a functionally distinct urgency signal, the next candidate variables should be persistence of exposure, recovery failure after offset, uncontrollability or inescapability, and prediction-error weighted threat. suggested_experiment_type: discriminative_pair - id: INV-065 title: proxy_goal_necessity claim_type: invariant status: candidate subject: goal.multi_step_planning description: > Any agent architecture with a bounded planning horizon and a sparse reward landscape requires intermediate wanting-rich states (proxy goals) to exhibit multi-step goal-directed behavior beyond the planning horizon. In REE terms: the CEM planner operates over a fixed horizon (~10 steps). In environments where resources are sparse or distal, resource z_world states will frequently lie outside that horizon. Without intermediate states that have acquired wanting value through schema binding or trajectory consolidation, the CEM score surface has no gradient toward the resource -- the agent cannot plan beyond 10 steps. Proxy goals are not a design choice but an architectural necessity: they are whatever z_world states have acquired wanting value through experience-dependent association, making the wanting landscape navigable at all scales. This is an invariant prediction about any planning architecture with these properties, independent of implementation details. notes: > The proxy goal mechanism emerges automatically in REE via two routes: (1) E1 schema-conditioned wanting (MECH-216): schema slots predicting resource encounter seed VALENCE_WANTING before direct contact, extending wanting coverage to E1-predictive z_world states. (2) Trajectory-consolidated wanting (MECH-217): reverse replay spreads wanting backward along approach paths, making every waypoint a plannable sub-goal. Neither route requires explicit proxy goal design -- they are consequences of the wanting landscape architecture (SD-014, SD-015) combined with E1 schema formation and offline consolidation. The invariant predicts that resource acquisition rate in sparse environments will be strongly gated by the density of proxy goals in the wanting landscape. registered: "2026-04-08" depends_on: [SD-014, SD-015, MECH-216, MECH-217] - id: MECH-216 title: e1_predictive_wanting claim_type: mechanism status: candidate subject: goal.e1_schema_wanting implementation_phase: v3 implementation_note: > IMPLEMENTED 2026-04-09. E1DeepPredictor.schema_readout_head (Linear(hidden_dim,1)+Sigmoid) reads LSTM top-layer hidden state -> schema_salience [0,1]. Agent caches in _schema_salience via _e1_tick(), seeds VALENCE_WANTING when > threshold via update_schema_wanting(). Config: E1Config.schema_wanting_enabled (False default), REEConfig.schema_wanting_threshold (0.3), schema_wanting_gain (0.5). Training: compute_schema_readout_loss(resource_proximity_target). Validation: EXQ-263 queued. description: > E1 schema activations that reliably predict resource encounters seed VALENCE_WANTING at the current z_world position before resource contact. This is a Pavlovian conditioned wanting mechanism: E1 acts as a CS-detector (firing on environmental features that precede resources), and its activation seeds the same wanting signal as the UCS (resource contact itself). The result is a wanting landscape that extends beyond directly-encountered resource positions -- it also includes z_world states from which resources are predictable via E1's learned schemas. The hippocampus then navigates toward these E1-predicted wanting regions, producing approach behavior toward schema-activating intermediate states (proto-proxy goals) before the direct wanting signal from SD-014 is available. notes: > Mechanistic basis: E1's LSTM hidden state encodes schema activations. These schemas form from temporal co-occurrence: E1 learns to predict z_world_t+1 from z_world_t, and internally differentiates episodes in which resource contact subsequently occurs from those that do not. Schema activation (the internal E1 state preceding resource-predictive transitions) becomes the CS signal. Implementation path in V3: requires reading out E1 schema activation magnitude at each waking step and using it to modulate SerotoninModule.update_benefit_salience() (which currently requires resource contact to update VALENCE_WANTING). A schema activation signal above threshold would seed partial wanting without contact. This is the forward-looking complement to MECH-217's backward consolidation. DR-13 (z_self temporal depth) is not required -- this operates in z_world domain. Distinguishable from direct contact wanting by spatial offset: E1 wanting fires at approach positions, contact wanting fires at resource position. registered: "2026-04-08" depends_on: [SD-014, SD-018, MECH-203, INV-065] - id: MECH-217 title: temporal_wanting_propagation claim_type: mechanism status: candidate subject: goal.replay_wanting_spread implementation_phase: v3 description: > Offline consolidation via reverse replay (MECH-165) spreads VALENCE_WANTING backward along previously successful approach trajectories. During SWS replay, the reversed world_state sequence of resource-approach episodes is replayed through the residue field. Each waypoint along the reversed trajectory receives a fractional wanting update proportional to the resource-contact wanting at the episode terminus and decayed by distance from the contact event. This produces a wanting gradient along the approach path: positions closer to the resource acquire stronger wanting signal than positions farther away, creating a navigable gradient from any waypoint. The practical effect is that multi-step goal-directed behavior emerges from single-resource encounters: the agent can navigate toward any previously-traversed waypoint (within CEM horizon) that was part of a resource-approach path. notes: > Mechanism requires MECH-165 (reverse replay substrate). MECH-165 IMPLEMENTED 2026-04-09 (exploration buffer + reverse replay + diverse scheduler). Substrate gap resolved. The wanting-spread function can be added to HippocampalModule.reverse_replay() -- after generating the reversed Trajectory, iterate through world_states and call ResidueField.update_valence() with decayed wanting at each step. Decay schedule: wanting_at_waypoint = wanting_at_terminus * gamma^(steps_from_terminus), gamma ~0.9. RBF bandwidth handles spatial smoothing. This is the consolidation-mediated complement to MECH-216's waking schema conditioning. Together they populate the wanting landscape at all temporal scales: MECH-216 (waking, forward-looking), MECH-217 (offline, backward-spreading). Distinguishable from MECH-216: MECH-217 wanting is path-specific (concentrated along previously traversed routes), MECH-216 wanting is schema-specific (concentrated at feature-activating positions regardless of prior path). registered: "2026-04-08" depends_on: [MECH-165, SD-014, INV-065] - id: MECH-218 title: interoceptive_predictive_wanting claim_type: mechanism status: candidate subject: goal.interoceptive_wanting implementation_phase: v4 description: > When z_self depletion trend is established (energy falling before drive_level becomes acutely high), E2's self-model projects forward and seeds anticipatory homeostatic wanting -- VALENCE_WANTING populated by predicted future deficit rather than by current resource proximity or schema activation. This produces approach behaviour that begins before the need is critical: seeking before thirsty, rather than seeking when thirsty. The signal flows from E2 self-transition predictions (depletion rate) through GoalConfig.drive_weight modulation into VALENCE_WANTING seeding, effectively increasing the drive-level multiplier before the actual drive_level threshold is crossed. This is the anticipatory homeostatic extension of SD-012 (reactive homeostatic drive modulation). notes: > V4 gated. Requires DR-13: z_self temporal depth. Without temporal depth in z_self (currently body_obs -> MLP -> EMA, no recurrence), E2 cannot estimate depletion *rate*, only current depletion *level*. DR-13 requires either recurrent z_self encoder or E2 feedback into z_self enrichment (E2 as self-model with multi-step prediction). Once DR-13 is available: the depletion rate signal can modulate SerotoninModule (or a parallel interoceptive neuromodulator) to inject wanting proportional to projected future deficit. Distinguishable from SD-012 (reactive): reactive wanting fires on current drive_level; predictive wanting fires on drive_level_rate (trend). Dissociation: an agent with a depleted but stable energy level should NOT show MECH-218 wanting; an agent with rapidly-falling energy at high baseline should show MECH-218 wanting before the absolute threshold is crossed. Clinical note: anticipatory homeostatic regulation failure (unable to seek before critical deficit) is a component of executive dysfunction in ADHD and depression where prospective self-model accuracy is degraded. registered: "2026-04-08" depends_on: [SD-012, MECH-203, INV-065] - id: ARC-051 title: multi_level_wanting_goal_hierarchy claim_type: architecture_hypothesis status: candidate subject: goal.emergent_hierarchy implementation_phase: v3 description: > The combination of the VALENCE_WANTING landscape (SD-014), E1 schema-conditioned wanting (MECH-216), and trajectory-consolidated wanting (MECH-217) produces an emergent multi-level goal hierarchy without explicit goal-hierarchy design. Level 1: direct resource encounter wanting (SD-018 + MECH-203, short-range, precise). Level 2: E1-schema-mediated predictive wanting (MECH-216, medium-range, schema- specific). Level 3: trajectory-consolidated wanting (MECH-217, long-range, path-specific). All three levels populate the same VALENCE_WANTING field; the CEM scorer responds to the combined gradient, enabling planning at all scales from a single unified navigation mechanism. The architecture is sufficient for multi-step goal-directed behavior without introducing a separate goal-planning subsystem. notes: > The hierarchy is emergent in the strict sense: it is not encoded anywhere as a hierarchy -- it is a consequence of three separate mechanisms (contact-wanting, schema-wanting, replay-wanting) that happen to operate at different spatial and temporal scales and all feed the same VALENCE_WANTING field. The CEM planner cannot distinguish which level it is responding to; it only sees the combined gradient. This has a prediction: removing any one level degrades multi-step performance at the scale that level operates -- removing MECH-217 degrades performance specifically in paths longer than the E1 schema radius; removing MECH-216 degrades performance in environments with reliable approach cues but no clear CEM-horizon gradient. The V4 interoceptive level (MECH-218) extends the hierarchy into the z_self domain (ARC-031 territory) but is not required for the V3 multi-step claim. Level 4 (MECH-218, V4): interoceptive anticipatory wanting (requires DR-13). registered: "2026-04-08" depends_on: [SD-014, MECH-216, MECH-217, INV-065, SD-018] - id: SD-020 title: "z_harm_a encodes affective surprise (precision-weighted PE), not raw accumulated harm state. The affective stream training target is the mismatch between predicted and actual harm accumulation, making E3 urgency respond to unexpectedness of threat, not just magnitude." claim_type: design_decision subject: harm_stream.affective_surprise_pe polarity: asserts status: candidate implementation_phase: v3 claim_level: architectural depends_on: [SD-011, SD-019, ARC-016] functional_restatement: > The current AffectiveHarmEncoder trains z_harm_a via an auxiliary head predicting raw accumulated harm scalar. This makes z_harm_a an integrator of absolute state, not a surprise detector. Chen (2023, Front Neural Circuits) establishes that the anterior insula cortex (AIC) -- the biological analog of z_harm_a -- responds to "unsigned intensity prediction errors as a modality-unspecific aversive surprise signal," not to raw stimulus magnitude. SD-020 requires replacing the aux loss target: instead of MSE(harm_accum_pred, accumulated_harm), train on MSE(harm_surprise_pred, harm_PE) where harm_PE = actual_accum - predicted_accum. This makes z_harm_a encode how SURPRISING the current threat level is, not how high it is. E3 urgency then responds to unexpected threat escalation (which requires immediate behavioral change) rather than sustained-but-expected threat (which the agent may already be handling). evidence_quality_note: > No experimental evidence yet. Derived from systematic comparison of Chen (2023) hierarchical predictive coding framework against current SD-011 implementation. Chen explicitly states AIC encodes PE, not raw state. Seymour (2019, Neuron) supports the framing: pain-as-RL-signal is a precision-weighted control signal, not raw nociceptive magnitude. Q-036 (already registered) asks exactly this question -- SD-020 is the "prediction error" answer to Q-036's open question. notes: > Addresses Q-036 directly. If SD-020 is confirmed experimentally, Q-036 answer is: "surprise (PE) is necessary in addition to temporal integration." The implementation requires a simple harm-accumulation predictor (e.g., linear model of harm_accum from previous steps) to generate the PE target. This is a low-complexity change to the training loop, not the encoder architecture. The encoder stays the same; only the aux loss target changes. registered: "2026-04-08" location: docs/architecture/sd_011_dual_nociceptive_streams.md - id: SD-021 title: "Descending pain modulation: when E3 is committed to a trajectory through expected harm, z_harm_s PE precision is reduced (endogenous analgesia). Commitment gates sensory harm attenuation." claim_type: design_decision subject: harm_stream.descending_modulation polarity: asserts status: candidate implementation_phase: v3 claim_level: architectural depends_on: [SD-011, SD-020, ARC-016, MECH-090] functional_restatement: > The harm stream is currently purely feed-forward: bottom-up encoding from observations to latent, consumed by E3. There is no descending pathway. Chen (2023) emphasizes that the pain system includes a descending inhibitory pathway (pgACC -> PAG -> RVM) that provides top-down suppression of nociceptive input during committed action. In REE terms: when E3 commits to a trajectory that passes through a hazard (accepted expected harm), the sensory harm stream z_harm_s should be precision-downweighted for the PREDICTED component of harm. The forward model E2_harm_s already predicts z_harm_s_next -- the precision of the residual (actual - predicted) should be reduced during commitment. Implementation: when E3.committed_trajectory is not None and beta_gate is elevated (MECH-090), multiply z_harm_s PE by a commitment-gated attenuation factor (0 < alpha_descending < 1). z_harm_a is NOT attenuated (affective load persists regardless of commitment -- you can tolerate expected pain but it still matters). evidence_quality_note: > No experimental evidence yet. Derived from Chen (2023) descending pathway emphasis and Keltner (2006) finding that expectation modulates sensory transmission through a distinct modulatory network. Keltner showed expectation activates caudal ACC/nCF (modulatory) while suppressing thalamus/S2/insula (sensory) -- this IS descending modulation. The REE implementation maps commitment -> expected harm prediction -> sensory PE attenuation. Biological grounding: endogenous opioid system provides analgesia during committed escape/approach behavior. notes: > This creates a commitment-dependent analgesic mechanism: an agent crossing a hazard to reach a goal suppresses sensory alarm for the expected harm component while maintaining affective urgency (z_harm_a still elevated). Disruption of this mechanism maps to hyperalgesia / anxiety disorders (inability to attenuate expected sensory harm during committed action). Connects to MECH-090 (beta gating) and MECH-094 (hypothesis tag). During beta-elevated commitment, the agent is executing, not planning -- sensory PE precision should be reduced. registered: "2026-04-08" location: docs/architecture/sd_011_dual_nociceptive_streams.md - id: MECH-220 title: harm_stream_hub_coordination claim_type: mechanism_hypothesis subject: harm_stream.cross_stream_hub polarity: asserts status: candidate implementation_phase: v3 depends_on: [SD-011, SD-020, SD-021, ARC-016] description: > A lightweight cross-stream coordination mechanism (harm hub) allows z_harm_s and z_harm_a to share context without collapsing into a single stream. The hub implements two directed interactions: (1) z_harm_a receives z_harm_s PE magnitude as context (sensory surprise informs affective load update rate); (2) z_harm_s precision is gated by z_harm_a urgency level (high affective load amplifies sensory sensitivity -- the hypervigilance pathway). This is NOT a fusion: the streams maintain separate encoders, separate training targets, and separate functional roles. The hub is a gating/attention mechanism that coordinates their interaction. notes: > Derived from Chen (2023) "cingulate-insula hub" concept. Chen's central claim: the two pain pathways are not fully independent parallel channels but are coordinated by a cingulate-insula hub with four functions: salience detection, cortical-subcortical integration, network switching, and ACC functional coupling. Our current implementation has z_harm_s and z_harm_a as completely independent encoders with zero cross-talk. The hub mechanism addresses this gap with minimal architectural change: a single attention/gating layer after the two encoders, parameterised by hub_weight. Testable prediction: hub_weight > 0 should improve both forward model R^2 (z_harm_s benefits from affective context for precision) and urgency calibration (z_harm_a benefits from sensory PE for surprise detection). A hub_weight ablation should show performance degradation vs the coordinated version. registered: "2026-04-08" location: docs/architecture/sd_011_dual_nociceptive_streams.md - id: ARC-052 title: harm_precision_weighting claim_type: architecture_hypothesis subject: harm_stream.dynamic_precision polarity: asserts status: candidate implementation_phase: v3 depends_on: [SD-011, SD-020, ARC-016, MECH-220] description: > Each harm stream outputs a precision estimate alongside its latent vector. E3's use of harm streams is weighted by their respective precisions: high-precision z_harm_s has more influence on attribution (SD-003), high-precision z_harm_a has more influence on commit gating. Precision is context-dependent: (1) z_harm_s precision increases with forward model accuracy (when E2_harm_s predictions are good, the PE is more informative); (2) z_harm_a precision increases with accumulation stability (when threat state is changing rapidly, z_harm_a is less precise -- high volatility reduces confidence in the accumulated state). Neuromodulatory inputs (serotonin, norepinephrine) modulate the precision gains, matching Chen (2023) ACh/NE/DA precision control framework. notes: > This extends ARC-016 (dynamic precision) into the harm stream specifically. Chen (2023) formalizes gain = xi_2 / (xi_2 + xi_1) where relative precision determines how much PE updates the prediction. Our current harm encoders have fixed architectures with no precision output. Adding a precision head (sigma output alongside mu) to each encoder enables the precision-weighted framework. Implementation: HarmEncoder and AffectiveHarmEncoder each output (z, log_sigma); E3 weights inputs by exp(-log_sigma). V3 scope: implement precision heads but use fixed precision initially; modulate in experiments. V4: wire neuromodulatory control (SerotoninModule already exists). registered: "2026-04-08" location: docs/architecture/sd_011_dual_nociceptive_streams.md