Roadmap

Claim Type: implementation_note Scope: Program phases, repository roles, and phase-gate criteria Depends On: IMPL-020, IMPL-021, IMPL-022, MECH-057, MECH-058, MECH-059, MECH-060 Status: candidate Claim ID: IMPL-008

---

Status Snapshot (2026-06-10T01:10Z – nightly docs sync; heavy substrate + governance + thought-intake day. Substrate landings since the 2026-06-09T01:10Z snapshot all sit on top of the in-flight 2026-06-09 substrate cluster: MECH-189 super-ordinal goal-anchor ContextMemory writes substrate (ree-v3 main c7ac035; REE_assembly master f212523d80 + docs/architecture/mech_189_super_ordinal_goal_anchors.md) closes infant_substrate:GAP-11. New SuperOrdinalGoalMemory in ree_core/goal.py (agent-owned, NOT reset per episode, cue-indexed key=z_world context / value=z_goal anchor); WRITE hook at agent.update_z_goal (child phase) gated on the MECH-189 conjunction (high-salience drive-modulated benefit) AND (high contextual complexity); complexity policy is pluggable via super_ordinal_complexity_mode (default novelty-vs-anchors). Recurring high-salience contexts REINFORCE their anchor on salience alone toward the matured z_goal. READ hook at the top of update_z_goal seeds adult z_goal via GoalState.cue_pull from the retrieved anchor when z_goal is below floor. 985/985 contracts + 8 new MECH-189 contracts + 7/7 preflight PASS; bit-identical OFF; no phased training; MECH-094 N/A. V3-EXQ-588c readiness diagnostic queued (LOAD-BEARING C1 = ADULT z_goal seeding DISCRIMINATION ARM_ON vs ARM_OFF). MECH-189 stays candidate (implementation_note only). MECH-294 multi-content theta-burst packet substrate + compose-coherence amend (ree-v3 main 53d4c75 + amend af97b23; REE_assembly master 4c680c03f5) wires per-cycle joint binding of {goal, action, risk, state} into a ThetaPacket. The amend (parameter-free, no trained head, bit-identical OFF) reads within-cycle co-binding via ThetaPacket.currency_coherence() (joint ~1.0 / alternation ~0.25 / shuffled 0.0) so the binding MODE causally reaches E3 behaviour rather than collapsing onto action-only similarity (the V3-EXQ-657a coherence-metric autopsy fork). New no-op flag theta_packet_compose_use_joint_coherence (default True; False = legacy action-only ablation). V3-EXQ-661 substrate-readiness 5-arm diagnostic queued (TV-margin readouts + non-vacuity gate; PASS routes to a separate behavioural-evidence successor session). MECH-294 stays candidate / v3_pending. SD-033b GAP-8 OFC trainable state_bias_head (ree-v3 main 382db2c + queue 8839724; REE_assembly master e7dc4f7152 + closure-plan node c5460ebb9f + SD-033b evidence_quality_note c8546ae0ff) is the SD-033a GAP-D mirror. OFCConfig.train_state_bias_head (default False -> last Linear zeroed, bit-identical OFF) + bias_head_parameters() + REEConfig ofc_train_state_bias_head (from_dims + agent.py OFC build-site getattr). When True the OFC state_bias_head last Linear keeps random init so it trains via E3-gradient REINFORCE, unblocking the deferred behavioural arm. Substrate constraint: OFC reads only z_world + z_harm (no appetitive / drive input) -> behavioural readout must be AVERSIVE devaluation (ofc_harm_dim > 0). V3-EXQ-485d 2-arm readiness diagnostic queued (head weight-delta-from-init load-bearing C2 under SP-CEM diversity). SD-033b stays candidate / v3_pending; this row only confirms the trainable arm is wired. ARC-063 cross-episode rule-persistence flag amend (ree-v3 main, 2026-06-09) advances V3-EXQ-654 GAP-B maturity for the ARC-063 distributed CandidateRule field. MECH-295 RE-SCOPED + goal_pipeline:GAP-4 CLOSED by governance decision (REE_assembly master) – modulatory reading retained; behavioural-necessity terminally FALSIFIED by V3-EXQ-490j severed-bridge baseline; the modulatory-SUFFICIENCY behavioural retest is preserved as OPTIONAL non-GAP-blocking. closure_status.md done 44 -> 45, remaining 24 -> 23, overall_progress 72.4% -> 73.2%. self_attribution GAP-1 / GAP-2 / GAP-3 + infant_substrate:GAP-13 re-adjudicated (REE_assembly master) – both gates collapse onto the SHARED selection-authority frontier paced by V3-EXQ-660 (MECH-341) + V3-EXQ-643a authority + V3-EXQ-604c, since ARC-065 SP-CEM became the main-path default 2026-05-17 AND the V3-EXQ-543l sleep-substrate run + V3-EXQ-614e CEM-pool collapse autopsy together prove the SP-CEM lever alone does not move the monomodal-equilibrium frontier. No code, no claim re-weighting; resume_conditions repointed. IGW generator suppresses V4+ scoped claims by PHASE/VERSION (REE_assembly master 3c3f37f2ff) – _proposed_experiments now skips implementation_phase: v4|v5 and version-band-v4+ claims independently of epistemic_category. Effect: removes 7 plain-v4 proposal leaks (SD-033e / MECH-264 / MECH-265 / MECH-129 / MECH-130); V4 claims are now invisible to the IGW proposal lane through either the substrate_conditional epistemic-category lever OR the phase / version lever. Heavy thought-intake + REAP cycle – six intake sessions registered V4 / post-V5 candidate clusters all status:candidate / epistemic_category:substrate_conditional / version-scoped: INV-077 evaluation-channel-integrity + Q-069 + EXT-008 meta-agent failure mode; ARC-087 / MECH-372 / Q-065 sense-specific perceptual manifolds; ARC-089 / MECH-374 / Q-066 spintronic-memristive cognifold substrate (post_v5); ARC-085 / MECH-365 / MECH-366 / Q-060 competitive-interactions cluster; MECH-359 / MECH-360 / MECH-361 candidate-differentiated affective gradients; MECH-362 + Q-057 CA3 developmental sparsification; MECH-364 + Q-059 laughter load-release. The user’s standing rule was reaffirmed twice in the cycle: genuinely-new intake content is REGISTERED into claims.yaml in the same pass that authors the home doc, NOT left as “future-registration” prose. infant_substrate:GAP-14 prereq-c InfantCurriculumScheduler Phase 0->1 advancement gate retune in flight under the active 21:48Z claim (queues V3-EXQ-591b readiness diagnostic). evidence/experiments/ flat top-level holds 346 v3exq* manifests (recursive count ~1053 incl. nested per-run dirs); per-machine fleet runner_status/*.json cards aggregate to 848 unique V3 queue_ids completed (286 PASS / 429 FAIL / 91 ERROR / 41 UNKNOWN / 1 INCONCLUSIVE; 1632 total pre-dedup); pending_review.md 0 items (V3-EXQ-603i precondition_unmet diagnostic was adjudicated and the autopsy routed successor work onto the 603-lineage behavioural bridge re-test programme); queue items[] 3 items, all claimed – V3-EXQ-603l SD-059 / MECH-358 escape-affordance bridge BEHAVIOURAL re-test (ree-cloud-2), V3-EXQ-655 INV-074 crystallization-necessity task-distribution-shift Phase-3 redesign (ree-cloud-1; supersedes V3-EXQ-610f), V3-EXQ-660 MECH-341 committed-class diversity RETEST on the GAP-A-ready / authority-ready stack using a WITHIN-CLASS-REPRESENTATIVE-DIVERSITY readout (DLAPTOP-4.local).)

• Substrate / governance landings since the 2026-06-09T01:10Z snapshot:
  • MECH-189 super-ordinal goal-anchor ContextMemory writes substrate (development.super_ordinal_goal_anchor_writes, ree-v3 main c7ac035; REE_assembly master f212523d80) – closes infant_substrate:GAP-11. New SuperOrdinalGoalMemory on the agent (cue-indexed key=z_world context / value=z_goal anchor; NOT reset per episode) with write hook at agent.update_z_goal (child phase) gated on the MECH-189 conjunction (high-salience drive-modulated benefit) AND (high contextual complexity) (complexity policy pluggable via super_ordinal_complexity_mode, default novelty-vs-anchors; recurring high-salience contexts REINFORCE their anchor on salience alone). Read hook seeds the adult z_goal via GoalState.cue_pull from the retrieved anchor when z_goal is below floor. 985/985 contracts + 8 new MECH-189 contracts + 7/7 preflight PASS; bit-identical OFF. V3-EXQ-588c readiness diagnostic queued (LOAD-BEARING C1 = ADULT z_goal seeding DISCRIMINATION ARM_ON vs ARM_OFF). MECH-189 stays candidate – implementation_note only.
  • MECH-294 multi-content theta-burst packet + compose-coherence amend (binding.theta_burst_packet, ree-v3 main 53d4c75 + af97b23; REE_assembly master 4c680c03f5) – per-cycle joint binding of {goal, action, risk, state} into a ThetaPacket; compose path reads within-cycle co-binding via currency_coherence() (joint ~1.0 / alternation ~0.25 / shuffled 0.0) so the binding MODE causally reaches E3 behaviour rather than collapsing onto action-only similarity. Parameter-free, no trained head, bit-identical OFF; new no-op flag theta_packet_compose_use_joint_coherence (default True; False = legacy action-only ablation); 8/8 packet contracts + 7/7 preflight PASS. V3-EXQ-661 substrate-readiness 5-arm diagnostic queued (TV-margin readouts across OFF / JOINT / ALTERNATION / SHUFFLED / ALT-coh-OFF; non-vacuity gate). MECH-294 stays candidate / v3_pending.
  • SD-033b GAP-8 OFC trainable state_bias_head (commitment_closure:GAP-8, ree-v3 main 382db2c + queue 8839724; REE_assembly master e7dc4f7152 + closure-plan node c5460ebb9f + SD-033b evidence_quality_note c8546ae0ff) – SD-033a GAP-D mirror. OFCConfig.train_state_bias_head (default False -> last Linear zeroed, bit-identical OFF) + bias_head_parameters() + REEConfig ofc_train_state_bias_head. When True the OFC state_bias_head last Linear keeps random init so it trains via E3-gradient REINFORCE, unblocking the deferred behavioural arm. Substrate constraint: OFC reads only z_world + z_harm (no appetitive / drive input) -> behavioural readout must be AVERSIVE devaluation. V3-EXQ-485d 2-arm readiness diagnostic queued (head weight-delta-from-init load-bearing C2 under SP-CEM diversity; PASS unblocks the full GAP-8 behavioural arm, NOT yet queued). SD-033b stays candidate / v3_pending.
  • ARC-063 AMEND: cross-episode rule-persistence flag (policy.rule_apprehension_layer.candidate_rule_field.cross_episode_persistence, ree-v3 main, 2026-06-09) – advance for V3-EXQ-654 GAP-B maturity. Allows minted candidate rules to PERSIST across episodes (off by default; opt-in flag) so the rule-creator’s eligibility credit pool is not zeroed on every episode reset.
  • MECH-295 RE-SCOPED + goal_pipeline:GAP-4 CLOSED (REE_assembly master, 2026-06-09) – user-approved governance decision reversing the 2026-06-04 “queue a V3-EXQ-490L successor” stance: 490k confirmed the modulatory-SUFFICIENCY behavioural retest CANNOT run on the current substrate (argmin_flip_ticks=0 by construction; collapsed candidate pool / weak-z_goal ceiling owned by scaffolded_sd054_onboarding). MECH-295 re-scoped to the MODULATORY reading (behavioural-necessity FALSIFIED by V3-EXQ-490j severed-bridge baseline; modulatory reading substrate-supported by 490j C6/C7/C9 + V3-EXQ-493 isolation); the modulatory-SUFFICIENCY retest is preserved as OPTIONAL non-GAP-blocking. closure_status.md overall_progress 72.4 -> 73.2%; MECH-295 stays status=candidate. NO promotion (substrate-firing support is not behavioural validation).
  • self_attribution GAP-1 / GAP-2 / GAP-3 + infant_substrate:GAP-13 re-adjudicated (REE_assembly master) – both gates’ resume_conditions were silently STALE: ARC-065 SP-CEM became the main-path default 2026-05-17 (the GAP-2 condition’s premise), and V3-EXQ-543l sleep-substrate + V3-EXQ-614e CEM-pool collapse autopsy together prove SP-CEM alone does not move the monomodal-equilibrium frontier. Both gates collapse onto the SHARED selection-authority frontier paced by V3-EXQ-660 + V3-EXQ-643a authority + V3-EXQ-604c. No code, no claim re-weighting; resume_conditions repointed; governance_2026_06_09 notes on both. The infant_substrate trio’s net (user’s “are those substrates not actionable?”): GAP-14 rides goal_pipeline:GAP-4 (now CLOSED) / the 603-line z_goal frontier; GAP-13 collapses into the shared authority frontier; GAP-11 (ContextMemory writes for MECH-189) is the only genuine separate substrate – the MECH-189 row above closes it.
  • IGW generator suppresses V4+ scoped claims from the proposal lane by PHASE/VERSION (REE_assembly master 3c3f37f2ff) – _proposed_experiments now skips implementation_phase: v4|v5 and version-band-v4+ claims (v3_v4 etc. kept). Removes 7 plain-v4 proposal leaks (SD-033e / MECH-264 / MECH-265 / MECH-129 / MECH-130); MECH-120 (v3) correctly retained. Effect: a V4 claim is invisible to the IGW proposal lane via EITHER lever (substrate_conditional epistemic category OR v4/v5 phase). Tooling-only; no claims.yaml change, no substrate, no experiment.
  • Heavy thought-intake + REAP cycle (REE_assembly master, 2026-06-09) – six intake sessions registered V4 / post-V5 candidate clusters all status:candidate / substrate_conditional / version-scoped (off the GAP-7 / V3 critical path). Genuinely-new content REGISTERED into claims.yaml in the same pass that authors the home doc (not left as “future-registration” prose – the user’s standing rule, reaffirmed twice today): INV-077 evaluation-channel-integrity + Q-069 + EXT-008 meta-agent failure mode (meta-development concern about the assembly loop itself); ARC-087 / MECH-372 / Q-065 sense-specific perceptual manifolds; ARC-089 / MECH-374 / Q-066 spintronic-memristive cognifold substrate (post_v5 physical-substrate compass); ARC-085 / MECH-365 / MECH-366 / Q-060 competitive-interactions cluster (cross-individual dynamics); MECH-359 / MECH-360 / MECH-361 candidate-differentiated affective gradients (V4/V5; the V3-narrow range-not-magnitude principle DELIBERATELY NOT registered, already owned by V3-EXQ-643a + behavioral_diversity_isolation:GAP-A/B + MECH-341 + SD-056); MECH-362 + Q-057 CA3 developmental sparsification (V4/V5); MECH-364 + Q-059 laughter load-release (V4; MECH-110 cross-references retained). All passed validate_claims --strict.
  • infant_substrate:GAP-14 prereq-c InfantCurriculumScheduler Phase 0->1 advancement gate retune – active 21:48Z claim covers a Phase 0->1 advancement-signal verification + V3-EXQ-591b substrate-readiness diagnostic queue (not yet landed at this snapshot; ree-v3 / REE_assembly edits pending).
• Runner activity since the 2026-06-09T01:10Z snapshot: per-machine runner_status/*.json cards continue to update across DLAPTOP-4.local + Daniel-PC + EWIN-PC + ree-cloud-1..4 + ree-worker-3. evidence/experiments/ flat top-level holds 346 v3exq* manifests on disk; recursive count incl. nested per-run dirs ~1053. Fleet aggregate 848 unique V3 queue_ids completed after dedup (286 PASS / 429 FAIL / 91 ERROR / 41 UNKNOWN / 1 INCONCLUSIVE; 1632 total pre-dedup).
• Queue depth: 1 -> 3 items at this read, all claimed – V3-EXQ-603l (priority 260, ree-cloud-2; SD-059 / MECH-358 escape-affordance bridge BEHAVIOURAL re-test, scored-evidence successor to V3-EXQ-603i, runs against the trainable-relief / safety-half head substrate landed 2026-06-09); V3-EXQ-655 (priority 250, ree-cloud-1; INV-074 crystallization-necessity task-distribution-shift Phase-3 redesign, supersedes V3-EXQ-610f); V3-EXQ-660 (priority 250, DLAPTOP-4.local; MECH-341 committed-class diversity RETEST on the GAP-A-ready / authority-ready stack using a WITHIN-CLASS-REPRESENTATIVE-DIVERSITY readout, arms candidate_summary_source=e2_world_forward + modulatory selection authority gain 0.5 + SP-CEM + MECH-341 + SD-056).
• Pending review: 1 -> 0 items at this read – V3-EXQ-603i diagnostic self-route precondition_unmet was adjudicated and the autopsy routed successor work onto the 603-lineage behavioural bridge re-test programme (V3-EXQ-603l).
• Bottleneck (continuation): the selection-authority frontier paced by V3-EXQ-660 + V3-EXQ-604c is the cross-cutting blocker for self_attribution / behavioral_diversity_isolation:GAP-B / arc_062:GAP-B / sd_037_axis_b after the 2026-06-09 re-adjudication folded those gates onto the shared diversity stack. The 603-lineage survival / escape-affordance evidence loop (V3-EXQ-603l) is the second in-flight gate. V3-EXQ-485d / V3-EXQ-588c / V3-EXQ-661 are the three new substrate-readiness diagnostics seeded by today’s substrate landings (SD-033b GAP-8 trainable OFC head + MECH-189 super-ordinal goal anchors + MECH-294 compose-coherence). The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding.

Immediate Work Queue (This Cycle, 2026-06-10)

1. V3-EXQ-603l outcome adjudication – ree-cloud-2; SD-059 / MECH-358 escape-affordance bridge BEHAVIOURAL re-test (scored-evidence successor to V3-EXQ-603i). PASS would route SD-058 / MECH-357 + SD-059 / MECH-358 toward candidate -> provisional on the post-trainable-relief/safety-head substrate; FAIL routes to /failure-autopsy on which half of the bridge (relief / safety / linker) is the missing piece.
2. V3-EXQ-660 outcome adjudication – DLAPTOP-4.local; MECH-341 committed-class diversity RETEST under the WITHIN-CLASS-REPRESENTATIVE-DIVERSITY readout (Learning #2 of the 614e autopsy: committed-class entropy is the WRONG matched axis for a within-class lever). PASS unblocks the next governance pass on MECH-341 + the modulatory-authority axis; FAIL routes back to the cross-cutting selection-authority autopsy programme that paces self_attribution / behavioral_diversity_isolation:GAP-B / arc_062:GAP-B / sd_037_axis_b.
3. V3-EXQ-655 outcome adjudication – ree-cloud-1; INV-074 crystallization-necessity task-distribution-shift Phase-3 redesign (supersedes V3-EXQ-610f). The next read for the plasticity-injection / EWC closure-prediction line.
4. V3-EXQ-485d outcome adjudication – 2-arm OFC trainable state_bias_head readiness diagnostic (frozen vs trainable; load-bearing C2 on head weight-delta-from-init under SP-CEM diversity). PASS unblocks the FULL behavioural GAP-8 arm (aversive-devaluation; the SD-033b candidate -> provisional decisive bit); FAIL routes to bias_scale clamp / head-init calibration.
5. V3-EXQ-588c outcome adjudication – MECH-189 super-ordinal goal-anchor ContextMemory writes substrate readiness (LOAD-BEARING C1 = ADULT z_goal seeding DISCRIMINATION ARM_ON vs ARM_OFF; the matured-anchor norm ~0.37 ceiling on the untrained-encoder readiness harness so a discrimination PASS with a 0.4 near-miss routes to a trained-encoder successor, NOT a failure).
6. V3-EXQ-661 outcome adjudication – MECH-294 compose-coherence 5-arm substrate-readiness (TV-margin readouts across OFF / JOINT / ALTERNATION / SHUFFLED / ALT-coh-OFF + non-vacuity gate). PASS routes to a SEPARATE behavioural-evidence MECH-294 successor session.
7. V3-EXQ-591b queue + readiness adjudication (in flight under the active 21:48Z claim) – infant_substrate:GAP-14 prereq-c InfantCurriculumScheduler Phase 0->1 advancement-signal gate retune. PASS clears the GAP-14 prereq-c lane; FAIL routes to a curriculum re-tune session.
8. MECH-353 / MECH-354 / MECH-355 follow-ups GATED (carried from 2026-06-06): V3 discriminative experiment for MECH-353 (blocked-action env), gated /queue-experiment plan for MECH-354 (cue-authority lever multiplies the wanting signal so the build stays gated on the 640b / 603l decisive bits), V4-social substrate design deferred for MECH-355. All carried.

Status Snapshot (2026-06-09T01:10Z – nightly docs sync; SD/MECH/ARC landings since the 2026-06-06T01:10Z snapshot all converge on the GAP-2 / GAP-7 survival / escape / cue-authority loop: MECH-314a Phase-2 amend + ARC-065 GAP-A shared cand_world_summaries amend (ree-v3 main 2026-06-07; routed by failure_autopsy_V3-EXQ-648 + failure_autopsy_V3-EXQ-614e) both re-source the per-candidate signal feeding the E3-side bias channels from the SD-056-trained e2.world_forward(z0, a_i) predictions (cross-candidate spread ~0.1147) instead of the collapsed proposer first-step z_world (spread <0.01 under monostrategy) – MECH-314a covers ONLY the curiosity channel + auto-augmentation _candidate_spread; ARC-065 GAP-A extends the identical fix to the SHARED cand_world_summaries consumed by lateral_pfc / ofc / mech295 / gated_policy / tonic_vigor; bit-identical OFF; new REEAgent._curiosity_candidate_summaries + REEAgent._candidate_world_summaries helpers; V3-EXQ-648a + V3-EXQ-649 substrate-readiness validations queued with cand_world_pairwise_dist readiness preconditions. scaffolded_sd054onboarding AMEND curriculum decomposition Stage-H** (ree-v3 main 2026-06-07; routed by failure_autopsy_V3-EXQ-603f confirming the goal-formation + ecological-seeding chain is SOUND but P1 couples goal-pipeline unfreeze + hazard wean simultaneously) inserts an isolated hazard-avoidance Stage-H between P0 and P1 so the agent acquires hazard navigation BEFORE P1 throws it at hazards; 85/85 contracts. **ControlVector logging telemetry rec-B** (ree-v3 main 2026-06-07) read-only default-OFF four-signal control telemetry making value / effort / opportunity-cost-of-time / vigor separately inspectable each E3 tick + EXPOSING the ARC-068-vs-MECH-320 collapse (opportunity cost + vigor both = w*v_t for the SAME MECH-320 v_t scalar); recommendation B logging only (causal first-class opportunity-cost split + four-axis controller deferred post-green-board). **SD-058 / MECH-357 instrumental-avoidance acquisition substrate landed** (ree-v3 main 2026-06-07; new ree_core/pfc/infralimbic_avoidance_gate.py) – the instrumental-ACQUISITION half of avoidance REE was missing per Moscarello & LeDoux 2013 (REE had Pavlovian/defensive REACTION via SD-035 amygdala + MECH-279 PAG freeze but no learned avoidance; a freeze-only substrate freezes instead of learning to avoid; ilPFC must SUPPRESS CeA-driven freezing). Three pieces (per-candidate ASSERT action-pathway score-bias + ilPFC freeze-suppression gate at MECH-279 application site + eligibility-trace avoidance_efficacy learning) all behind use_instrumental_avoidance (False default); PROTECTIVE-SCAFFOLD anneal in Stage-H (effective_efficacy = max(efficacy, scaffold_floor) with scaffold_avoidance_scaffold_floor_start=0.8/..._floor_end=0.0 – maternal-buffering / Turchetta 2020). LOAD-BEARING PREREQUISITE FOUND: legacy scaffold called sense() with NO harm args so z_harm_a was None across the entire curriculum leaving MECH-279 + SD-035 + SD-058 ALL inert; new scaffold_feed_harm_stream flag feeds env harm_obs + harm_obs_a so PAG/SD-035/SD-058 actually see threat (~0.34 in Stage-H). 912 contracts; v3_pending. V3-EXQ-603h FAILed adjudicated engaged-but-insufficient (gate engaged + suppressed PAG freeze on all INTACT seeds, readiness_met=true, but G_H_INTACT 0/3 not > LESION; scalar avoidance_efficacy decoupled from survival – seed-43 inversion); failure_autopsy_V3-EXQ-603h_2026-06-08 confirmed clean adjudicated; SD-058/MECH-357 stay candidate / v3_pending UNWEAKENED (claim_ids=[]); discovered dependency = relief/safety escape-affordance bridge. **SD-059 / MECH-358 relief/safety escape-affordance bridge substrate landed** (ree-v3 main 2026-06-08; new ree_core/pfc/escape_affordance_bridge.py) closes the 603h directed-escape gap. SD-058 suppressed the MECH-279 freeze but avoidance_efficacy is a GLOBAL SCALAR only penalising the no-op class – compute_action_bias by design does NOT compute the escape direction. Per Moscarello & LeDoux active avoidance also needs the LA/BA->NAcc relief/safety action-credit half; REE owned relief (MECH-302/SD-050) + safety (MECH-303/304/SD-052/SD-051) but they were UNWIRED to avoidance. Extends MECH-357’s scalar avoidance_efficacy into a per-FIRST-ACTION-CLASS credit table (relief half: directed action under threat that DROPS z_harm_a credits relief_affordance[action_class]; safety half: directed action after which threat is absent credits safety_affordance[action_class]); approach bonus under FUTURE threat = per-candidate NEGATIVE (favoured) score-bias toward each candidate whose first-action class carries combined affordance credit (no-op/freeze class never gets a bonus); three guards (bias_scale clamp + threat-context gate exactly zero when safe + per-tick leak – no pathological habit loop). DISTINCT from reflexive SD-037/MECH-281 escape (threat/arousal reflex) and from generic MECH-302/303/304 rows (fire on current state); this is learned-efficacy-gated DIRECTED approach binding an action to relief/safety for future-threat use. Bit-identical OFF; 8 new contracts. v3_pending until 4-arm validation EXQ PASSes. V3-EXQ-603i (ARM_BASE_IA_ONLY / ARM_RELIEF_BRIDGE / ARM_SAFETY_BRIDGE / ARM_RELIEF_SAFETY_BRIDGE) FAILed precondition_unmet (self-routed substrate_not_ready_requeue flagged for adjudication); substrate_queue CREATE escape-affordance-bridge (priority 1) carried over to the post-603i successor scaffolds. **Three post-603i successor scaffolds** (ree-v3 main 2026-06-08): (a) trainable escape-affordance learner module 7a0a417 (local PyTorch relief/safety heads + shared trunk + AdamW; continuous relief targets + response-produced safety targets + extinction targets; 23 contracts); (b) trainable relief/safety heads upgrade 58535af (lazy model + AdamW; detached compact state/context + action embedding; optimizer/no-op/simulation/hypothesis guards; prediction-based bounded threat-gated bias; 27 contracts); (c) E2 escape-affordance linker 6c856a5 (per user mid-session correction: REUSE/READOUT over E2 (cerebellar-analog) E2.world_forward – NOT a duplicate forward predictor; reads DETACHED action-consequence features for the executed (prev_z_world, action) pair into viability readouts harm_delta / threat_termination / safety_transition / refuge_reachability / survival_step + hippocampal-style per-action viability index + bounded threat-gated E3 bias behind own flag; 949 contracts + 7 preflight + boot matrix PASS). ALL THREE NOT validated substrate AND NOT changes to V3-EXQ-603i. **V3-EXQ-653 E2 escape-affordance linker readiness microdiagnostic** queued via /queue-experiment (ree-v3 main 2026-06-08T21:17Z) as the validation gate – forced-choice 4-action probe over the linker driven directly with controlled E2 consequence proxies + known outcomes; 4 arms (DISABLED_CONTROL / LINKER_READOUT_ONLY / LINKER_E3_BIAS / LINKER_TO_RELIEF_SAFETY_HEADS) x 3 seeds (42/43/44); readiness gates G0-G8; PASS routes back to queue a 603-lineage full behavioural bridge re-test; FAIL routes to failure-autopsy on this readiness diagnostic. **/governance cycle 2026-06-08T21:51Z applied substrate-side dispositions** (REE_assembly master d1efafaf31, 797-file regen) – carried 603i autopsy artifact 3c714ae49d (analysis-only; NO claims.yaml edits this cycle, both pending diagnostics claim-free); 603h substrate-side eq-note + source_autopsy on flat + runpack manifests (direction stays non_contributory); substrate_queue CREATE escape-affordance-bridge (priority 1); claims.yaml pending_retest_after_substrate on the survival-leg cohort ARC-060 / MECH-320 / ARC-068 / SD-054 (unchanged at this nightly read). **Runner ETA fix** (ree-v3 b559b5a, 2026-06-07) – multi-stage experiments without PASS/FAIL per-cell verdict lines now use the stable median-per-cell ETA path (640b “all over the place” symptom fixed; takes effect on next runner restart). evidence/experiments/ flat top-level holds **334 v3_exq* manifests (recursive ~994 incl. nested per-run dirs); per-machine fleet runner_status/*.json carries cards for DLAPTOP-4.local + Daniel-PC + EWIN-PC + ree-cloud-1..4 + ree-worker-3; pending_review.md 1 item (V3-EXQ-603i FAIL precondition_unmet flagged for adjudication); queue items[] 1 item (V3-EXQ-640b SD-057/MECH-346/MECH-347 cue-authority lineage CLEAN EVIDENCE retest, claimed DLAPTOP-4.local 2026-06-08T06:46Z; promotes 640a -> EVIDENCE on the post-643a authority gate))

• Substrate / governance landings since the 2026-06-06T01:10Z snapshot:
  • MECH-314a Phase-2 amend (e2.world_forward novelty-candidate-source) (policy.structured_curiosity_bonus.e2_world_forward_novelty_source, ree-v3 main) – routed by failure_autopsy_V3-EXQ-648_2026-06-07 (precondition_unmet; recommended_substrate_queue_entry.action=amend on MECH-314a-Phase-2-impl). Root cause: the MECH-314a per-candidate novelty AND the auto-augmentation _candidate_spread keyed on trajectory.world_states[:,0,:] (proposer first-step z_world) whose cross-candidate spread is <0.01 under monostrategy -> curiosity_bias_range=0.0 and curiosity_std_across_K=0.0 in EVERY arm including the ARM_1 positive control. SD-056-trained e2.world_forward(z0, a_i) predictions carry spread ~0.1147 (the representation the SD-056 readiness gate already validates) but were NOT the consumed signal – 648 readiness precondition measured e2.cand_world_pairwise_dist (0.1147 PASS) while C2 routed on the proposer-derived bias (<0.01) -> false READY -> self-route mislabelled a collapsed-input artefact as a wiring null (canonical V3-EXQ-642 same-statistic pattern). Fix: new REEAgent._curiosity_candidate_summaries(candidates) helper consulted FIRST in the curiosity block; when curiosity_candidate_source="e2_world_forward" (Literal, default “proposer” = legacy bit-identical), cur_summaries = e2.world_forward(z0.expand(K,-1), first_actions_K) [K, world_dim] – same construction the SD-056 cand_world_pairwise_dist diagnostic uses; both 314a RBF novelty AND _candidate_spread now key on the action-divergent representation. 877/877 contracts + 7/7 preflight PASS; 4 new C6 contracts. MECH-094 preserved (no_grad on the waking path). V3-EXQ-648a substrate-readiness validation queued with cand_world_pairwise_dist readiness precondition + the 604a curiosity non-degeneracy guard retained.
  • ARC-065 GAP-A: shared cand_world_summaries e2.world_forward source (policy.candidate_pool_per_candidate_signal_preservation.shared_channel, ree-v3 main) – routed by failure_autopsy_V3-EXQ-614e_2026-06-07. The shared-channel sibling of the MECH-314a Phase-2 curiosity amend landed earlier the same day: that pass fixed ONLY the curiosity channel’s consumed representation; this pass extends the identical e2.world_forward re-sourcing to the SHARED per-candidate cand_world_summaries consumed by ALL the other E3-side bias channels (lateral_pfc / ofc / mech295 / gated_policy / tonic_vigor). 614e signature: with the modulatory-bias-selection-authority gate now PROVEN operative (V3-EXQ-643a PASS), committed-class diversity still showed no lift (committed_class_entropy byte-identical across within-class T=0.5/1.0/2.0). Autopsy relocated the bottleneck from the authority gate (GAP-B, resolved) to the UPSTREAM candidate pool (GAP-A): all K CEM candidates produce identical z_world after one E2 world-forward step (cand_world_pairwise_dist=0.0000) despite differing first actions, so every E3-side bias channel sees a class-uniform pool. New REEAgent._candidate_world_summaries(candidates) helper (shared-channel sibling of _curiosity_candidate_summaries) consulted FIRST at all FIVE cand_world_summaries fresh-build sites (gated_policy block, lateral_pfc fallback, ofc fallback, mech295 fallback, and via reuse-chain the tonic_vigor anchor). Config: REEConfig.candidate_summary_source: Literal["proposer","e2_world_forward"] = "proposer" (default; bit-identical). Kept SEPARATE from curiosity_candidate_source (in-flight 648a validation) so the two compose without perturbing each other. 889/889 contracts (883 prior + 6 new) + 7/7 preflight PASS. V3-EXQ-649 substrate-readiness diagnostic queued. PASS unblocks the MECH-341 committed-class diversity re-test (the within-class-REPRESENTATIVE-diversity readout, NOT committed-class entropy per Learning #2 of the 614e autopsy).
  • scaffolded_sd054_onboarding AMEND curriculum decomposition / isolated Stage-H (curriculum.scaffolded_sd054_onboarding.curriculum_decomposition, ree-v3 main) – routed by failure_autopsy_V3-EXQ-603f_2026-06-07 (substrate-readiness FAIL, self-route substrate_not_engaged / foraging_competence_open, confirmed). 603f PROVED the goal-formation + ecological-seeding chain is SOUND – seed 44 foraged (P2 contact_rate 0.393, 85 events) AND seeded z_goal ecologically (z_goal_norm_at_contact_peak 0.450 > 0.4). The remaining GAP-2 blocker is the P1 SURVIVAL leg (G1 0/3; median episode len 12.5/38.0/28.5 vs gate 75; even the foraging seed 44 died at 28.5). ROOT CAUSE: P1 couples TWO competencies at once (goal-pipeline unfreeze + wean into the hazard band) and the agent cannot acquire both simultaneously; P0 trains only in the safe reef refuge so the agent never learns hazard navigation before P1 throws it at hazards. The fix (user-directed Stage-H only): a SEPARATELY-TRAINED isolated hazard-avoidance Stage-H between P0 and P1. New run_hazard_avoidance + HazardAvoidanceResult dataclass; goal pipeline FROZEN (seed_goal=False -> update_z_goal never called, z_goal untouched – the isolation); trains E1+E2 in a hazards-with-randomly-drifting-no-food-attraction env at midline spawn; measures median episode length over the last stability_window episodes vs survival_gate_steps (G_H survival readout, DIAGNOSTIC ONLY – does NOT abort the curriculum). New phase “hazard” in _build_env. Curriculum becomes Stage-0 -> Stage-0b -> P0 -> Stage-H -> P1 -> P2. All behind scaffold_hazard_stage_enabled (False default; bit-identical OFF). 85/85 scaffold contracts (79 prior + 6 new C12) + 7/7 preflight PASS. V3-EXQ-603g substrate-readiness validation (copy of 603f with Stage-H inserted) routed downstream to SD-058/MECH-357 + SD-059/MECH-358 below.
  • ControlVector logging (rec-B telemetry) (telemetry.control_vector_logging, ree-v3 main) – read-only default-OFF four-signal control telemetry making value / effort / opportunity-cost-of-time / vigor separately inspectable each E3 tick. EXPOSES the ARC-068-vs-MECH-320 collapse (opportunity cost AND vigor are both w*v_t for the SAME MECH-320 v_t scalar – ARC-068 is registered but unbuilt; its lit-pull is pending). Recommendation B (logging only); causal first-class opportunity-cost split (C) + full four-axis controller (D) DEFERRED post-green-board, gated on the ARC-068 lit-pull and MECH-320 regaining selection authority (V3-EXQ-643a / SD-056 / ARC-065 GAP-A). Modules: REEConfig.use_control_vector_logging (default False); bundle gains control_required + effort_term; E3Selector stores last_raw_scores; REEAgent _assemble_control_vector() writes _last_control_vector after e3.select. Schema includes V_outcome / C_effort / C_time / G_vigor / shared (logging tonic_vigor_v_t + w_action + w_passive so C_time.potential and G_vigor.potential are both computable as w*v_t for ONE scalar – the collapse made inspectable) / authority. Bit-identical OFF (contract C4); 889 contracts + 7 preflight + 4 new ControlVector contracts; activation smoke v_t=0.5 (forced floor; v_raw=-1.75 – the documented EXQ-624a sign/scale issue now visible in telemetry); C_time=G_vigor=0.05 (one scalar, two weights). Stage-B C_time<->G_vigor collapse-correlation diagnostic queued via /queue-experiment (claim_ids=[]; pre-registered rho ~ 1.0).
  • SD-058 / MECH-357 instrumental-avoidance acquisition (ilPFC-analog freeze-suppression + avoidance action pathway) (defensive_action.instrumental_avoidance_acquisition, ree-v3 main) – closes the V3-EXQ-603g G_H 0/3 + goal_pipeline:GAP-2 P1 survival-leg gap. Routed by failure_autopsy_V3-EXQ-603g-624c-651a_2026-06-07 + lit verdict targeted_review_hazard_avoidance_learning/SYNTHESIS.md (SD-035 x3 + MECH-279 + SD-054): the fix is STRUCTURAL not budgetary. REE had the Pavlovian/defensive REACTION side (SD-035 amygdala salience + MECH-279 PAG freeze) but lacked the instrumental-ACQUISITION side. Per Moscarello & LeDoux 2013, active avoidance is the resolution of a Pavlovian-instrumental conflict requiring the infralimbic PFC to SUPPRESS CeA-driven freezing (ilPFC lesion -> more freezing, less avoidance); a freeze-only substrate freezes instead of learning to avoid – the 603g G_H 0/3 signature. New ree_core/pfc/infralimbic_avoidance_gate.py (InstrumentalAvoidanceGate + InstrumentalAvoidanceGateConfig + InstrumentalAvoidanceGateOutput); pure-arithmetic regulator (no nn.Module, no trained params, no gradient flow); sibling to SD-035 CeA/BLA + MECH-279 PAG + MECH-313 NoiseFloor + MECH-320 TonicVigor + MECH-342 maintenance-release pattern. Three pieces all behind use_instrumental_avoidance (False default; bit-identical OFF): (a) instrumental-avoidance ACTION pathway – per-candidate E3 score-bias composed last in dacc_score_bias (after MECH-320 + MECH-353 assert) that under retained threat (z_harm_a) PENALISES the no-op/freeze class proportional to effective_efficacy * threat_scale (releasing the instrumental action; does NOT compute the escape direction – E3’s harm gradient ranks the directed candidates; ARC-007-strict-compatible); (b) ilPFC FREEZE-SUPPRESSION gate at the MECH-279 application site – when freeze_active AND should_suppress_freeze(z_harm_a_norm) (effective_efficacy * threat_scale >= suppression_threshold), the no-op override is SKIPPED so the agent takes its selected instrumental action; inert when use_pag_freeze_gate=False (the action-pathway half still operates – freezing is a passive no-op the bias penalises); (c) AVOIDANCE-EFFICACY LEARNING (eligibility trace) – a scalar avoidance_efficacy in [0,1] starting 0.0 (freeze-default); update in sense() (one-tick lag): directed action under threat that DROPS z_harm_a credits efficacy (EMA toward 1); freezing/failed-avoidance under threat decays it. PERSISTS across episodes within a stage; the gradual developmental acquisition (Debiec & Sullivan 2017 / Thompson 2008). PROTECTIVE-SCAFFOLD anneal: effective_efficacy = max(efficacy, scaffold_floor); Stage-H curriculum sets high floor early and anneals it as the learned efficacy takes over (maternal-buffering / Turchetta 2020 reset-curriculum analogue). LOAD-BEARING PREREQUISITE FOUND 2026-06-07: the legacy scaffold called sense(body, world) with NO harm args, so z_harm_a was None across the WHOLE curriculum – leaving MECH-279 PAG, SD-035 amygdala AND the SD-058/MECH-357 gate all INERT (they key on z_harm_a). New scaffold flag scaffold_feed_harm_stream (False default -> bit-identical) + module helper _sense_with_optional_harm feed the env harm_obs + harm_obs_a into sense() so z_harm_a is populated (~0.34 in Stage-H). The avoidance-driver experiments set it True; without it the gate has no threat signal to learn from. DISTINCT from the reflexive escape-from-freeze levers (do NOT read as a duplicate): SD-037 override_signal (orexin) raises PAG exit threshold and MECH-281 lowers MECH-091 urgency-interrupt – BOTH reflexive threat/arousal-driven escape. MECH-357’s ilPFC suppression is gated by LEARNED avoidance-efficacy bootstrapped by the protective-scaffold floor – the acquisition mechanism (Moscarello & LeDoux), not a reflex. 912 contracts + 7/7 preflight PASS; 7 new contracts in test_mech_357_instrumental_avoidance.py + 4 C13 in test_scaffolded_sd054_onboarding.py. v3_pending. V3-EXQ-603h FAILed adjudicated engaged-but-insufficient (gate engaged + suppressed PAG freeze on all INTACT seeds, readiness_met=true, but G_H_INTACT 0/3 not > LESION; scalar avoidance_efficacy decoupled from survival – seed-43 inversion: efficacy 0.633 -> worst survival). Failure_autopsy_V3-EXQ-603h_2026-06-08 confirmed AND clean adjudicated; SD-058/MECH-357 stay candidate / v3_pending UNWEAKENED (claim_ids=[]); discovered dependency = relief/safety escape-affordance bridge (SD-059 below).
  • SD-059 / MECH-358 escape-affordance bridge (relief/safety -> directed escape) (defensive_action.escape_affordance_bridge, ree-v3 main) – closes the V3-EXQ-603h directed-escape gap. SD-058/MECH-357 suppressed the MECH-279 freeze but avoidance_efficacy is a GLOBAL SCALAR that only penalises the no-op/freeze class – compute_action_bias by design “does NOT compute the escape direction”. 603h (engaged-but-insufficient, readiness met): gate suppressed freeze on all INTACT seeds but G_H_INTACT=0/3; seed-43 reached scalar efficacy 0.633 and survived WORST (11.0). The agent un-froze without acquiring a DIRECTED escape. Moscarello & LeDoux 2013: active avoidance ALSO needs the LA/BA->NAcc relief/safety action-credit half, not only suppression. REE owned relief (MECH-302/SD-050) + safety (MECH-303/304/SD-052/SD-051) but they were UNWIRED to avoidance – this is the wiring. New ree_core/pfc/escape_affordance_bridge.py (EscapeAffordanceBridge + Config + Output); pure-arithmetic, sibling to SD-058 in ree_core/pfc/. Extends MECH-357’s scalar avoidance_efficacy into a per-FIRST-ACTION-CLASS credit table (the minimal V3 rendering of escape_affordance[action] – directed escape direction in the discrete action space; location/policy indexing deferred). Two independently-toggleable halves (so the 4-arm validation dissociates): RELIEF half (MECH-302-consistent): directed action under threat that DROPS z_harm_a (delta = prev - now > relief_reward_floor) credits relief_affordance[action_class] EMA toward 1 – the d(z_harm_a)/dt<0 signal attributed to the specific action. SAFETY half (MECH-303/304-consistent): directed action after which threat is absent (threat_scale <= 0) credits safety_affordance[action_class] (response-produced safety / conditioned inhibition). Approach bonus (the directed escape): under FUTURE threat (threat_scale > 0), E3 receives a per-candidate NEGATIVE (favoured) score-bias toward each candidate whose first-action class carries combined affordance credit; the no-op/freeze class never gets a bonus. THREE guards: bias_scale clamp (cannot dominate the chain), threat-context gate (exactly zero when safe – never swamps food/goal approach), per-tick leak (forgetting – no pathological habit loop). DISTINCT from reflexive escape (SD-037 orexin / MECH-281 urgency – threat/arousal reflexes) and from the generic relief/safety rows (MECH-302/303/304 fire on the CURRENT state); this is learned-efficacy-gated DIRECTED approach binding an action to relief/safety for future-threat use. Bit-identical OFF (use_escape_affordance_bridge=False); 8 new SD-059 contracts in test_sd_059_escape_affordance_bridge.py PASS. MECH-094 standard simulation_mode no-op pattern (replay/DMN must not credit escape affordances or bias action selection on imagined outcomes). v3_pending until the 4-arm validation EXQ PASSes. V3-EXQ-603i substrate-readiness diagnostic (claim_ids=[]; thought-intake 2026-06-07 4-arm: ARM_BASE_IA_ONLY / ARM_RELIEF_BRIDGE / ARM_SAFETY_BRIDGE / ARM_RELIEF_SAFETY_BRIDGE + a nav-competence positive control so a flat G_H across all bridge arms is attributable to a survival/navigation ceiling rather than the bridge; non-vacuity gate: each enabled bridge half must increment its credit before G_H is scored) FAILed precondition_unmet – self-route flagged for adjudication via /failure-autopsy; substrate_queue CREATE escape-affordance-bridge (priority 1) carried over to the post-603i successor scaffolds below.
  • Three post-603i successor scaffolds (ree-v3 main 2026-06-08; NOT validated substrate; NOT changes to V3-EXQ-603i) – forward-looking experiments inspired by the 603i FAIL, scaffolded behind feature flags as successor OPTIONS not replacements for the active SD-059/MECH-358 arithmetic bridge. (a) Trainable escape-affordance learner (trainable_escape_affordance_learner, ree-v3 main 7a0a417 + 58535af): local PyTorch relief/safety heads (shared trunk + action embedding + AdamW) trained on continuous relief targets, response-produced safety targets, extinction targets. Off by default (use_trainable_escape_affordance_learner); 23 contracts on first pass + 27 contracts on the heads-trainable upgrade pass. (b) E2 escape-affordance linker (e2_escape_affordance_linker.py, ree-v3 main 6c856a5) – BUILT THE POST-603i SUBSTRATE AS A REUSE/LINKAGE READOUT OVER THE EXISTING E2 (cerebellar-analog) FORWARD MODEL per the user’s mid-session correction (NOT the originally-briefed escape_hazard_affordance_encoder, which would have duplicated E2). New E2EscapeAffordanceLinker (+ Config + Output): reads the DETACHED E2.world_forward feature for the executed (prev_z_world, action) pair into escape-affordance viability readouts (harm_delta / threat_termination / safety_transition / refuge_reachability / survival_step); exposes escape_affordance_features for the relief/safety heads; keeps a hippocampal-style per-action viability index (readout only – no trajectory gen / reward / selection); emits a bounded threat-gated E3 bias behind use_e2_escape_linker_e3_bias. TrainableEscapeAffordanceLearner gained optional extra_features (default None -> bit-identical). 11 contracts (10 guarantees + forced-choice readiness microdiagnostic 2/3-seed gates + agent E2-reuse); 949 contracts + 7 preflight + boot matrix PASS; py_compile PASS; validate_queue PASS. OFF by default; no backprop into E1/E2/E3; no sim/hypothesis learning; no no-op credit; relief/safety distinct. SD-059/MECH-358 (and MECH-302/303/304) NOT marked validated or weakened by this scaffold; no governance, no queue change, no 603j re-run queued. V3-EXQ-653 E2 escape-affordance linker readiness microdiagnostic queued via /queue-experiment (ree-v3 main 2026-06-08T21:17Z; claim_ids=[]; experiment_purpose=diagnostic) – forced-choice 4-action probe over the linker driven directly with controlled E2 consequence proxies + known outcomes; 4 arms (DISABLED_CONTROL / LINKER_READOUT_ONLY / LINKER_E3_BIAS / LINKER_TO_RELIEF_SAFETY_HEADS) x 3 seeds (42/43/44); readiness gates G0-G8. Does NOT validate/weaken SD-059/MECH-358 or MECH-302/303/304; not 603 closure. PASS routes back to queue a 603-lineage full behavioural bridge re-test; FAIL routes to /failure-autopsy on this readiness diagnostic.
  • Runner ETA fix (ree-v3 b559b5a, 2026-06-07) – multi-stage experiments without PASS/FAIL per-cell verdict lines now use the stable median-per-cell ETA path (Seed/Condition transitions counted as cell-completion boundaries + fallback pct guard in experiment_runner.py) instead of the divergent live-extrapolation fallback. 640b “all over the place” symptom fixed. Runner-code-only, no running experiment touched; takes effect on next runner (re)start.
  • /governance cycle 2026-06-08T21:51Z full regen (REE_assembly master d1efafaf31, 797-file regen) – carried + landed 603i autopsy 3c714ae49d; rebased onto origin’s 3 commits incl phase3-heartbeat; both verified on origin. NO claims.yaml edits this cycle (both pending diagnostics claim-free). 603h substrate-side eq-note + source_autopsy on flat + runpack manifests (direction stays non_contributory); substrate_queue CREATE escape-affordance-bridge (priority 1, ready=false); claims.yaml pending_retest_after_substrate on survival-leg cohort ARC-060 / MECH-320 / ARC-068 / SD-054 (cleared in prior governance cycle 2026-06-08T07:21Z apply pass; unchanged at this nightly read). The /governance command was run from the Mac via /sync after the Cowork sandbox could not commit (stale REE_assembly/.git/index.lock left by a crashed auto-gc the 797-file git add tripped; confirmed no holder process; cleared and pushed clean).
• Runner activity since the 2026-06-06T01:10Z snapshot: per-machine runner_status/*.json cards continue to update across DLAPTOP-4.local + Daniel-PC + EWIN-PC + ree-cloud-1..4 + ree-worker-3 (central-runner_status decoupled into per-machine files under evidence/experiments/runner_status/ post 2026-05-29 Phase 3 cutover). evidence/experiments/ flat top-level holds 334 v3exq* manifests on disk; recursive count incl. nested per-run dirs ~994.
• Queue depth: 3 -> 1 item at this read – V3-EXQ-640b SD-057 / MECH-346 / MECH-347 cue-authority lineage CLEAN EVIDENCE retest (claimed DLAPTOP-4.local 2026-06-08T06:46Z; re-runs the 640a scaffold cue-recall gain sweep – cue_recall_gain{0.2,1.0,5.0} x kappa{2.0,10.0} + ARM_OFF, 3 seeds – now that modulatory-bias-selection-authority is VALIDATED via V3-EXQ-643a PASS; use_modulatory_selection_authority=True+gain=0.5 on every arm; promotes 640a diagnostic -> EVIDENCE; supersedes V3-EXQ-640a; PRIMARY acceptance C_LIFT_PRIMARY = post_cue_approach_lift>0 over within-run background on >=2/3 seeds in the pre-registered decisive cell ARM_CUE_g5_k10). V3-EXQ-653 E2 escape-affordance linker readiness microdiagnostic (chip 22:17Z) being prepared by the runner.
• Pending review: 0 -> 1 item at this read – V3-EXQ-603i SD-059/MECH-358 escape-affordance bridge validation FAIL with self-route label substrate_not_ready_requeue flagged precondition_unmet. The diagnostic adjudication gate: the self-route’s premise did not hold, so the label must NOT drive a governance action (clear v3_pending / mint-or-AMEND substrate_queue / close-or-route a thought-intake) until adjudicated via /failure-autopsy. The manifest’s interpretation is the source of truth – a re-queued successor supersedes it.
• Bottleneck (continuation from yesterday’s framing): the GAP-2 / GAP-7 ecological-evidence loop remains primary. SD-058 + SD-059 are the latest substrate-side answers to the 603-lineage survival / escape leg, structuring the missing instrumental-avoidance ACQUISITION half (SD-058) + the directed escape via relief/safety affordance binding (SD-059). The in-flight gates: (a) V3-EXQ-603i adjudication via /failure-autopsy on the precondition_unmet flag – adjudicates whether the readiness gate’s premise actually held; if it did, the FAIL routes to a substrate revisit on the affordance-bridge half; if the precondition failed, the substrate stays held and the V3-EXQ-653 E2 escape-affordance linker readiness microdiagnostic is the next adjudicating bit. (b) V3-EXQ-653 – forced-choice 4-action probe over the post-603i E2 linker scaffold (claim-free diagnostic); PASS routes back to a full 603-lineage behavioural bridge re-test; FAIL routes to /failure-autopsy on this readiness diagnostic. (c) V3-EXQ-640b – promotes the 640a cue-authority diagnostic to EVIDENCE on the post-643a authority gate; PRIMARY acceptance is post-cue approach lift in the decisive cell. The three pending validations (648a curiosity + 649 shared / 603i + 653 / 640b) are the adjudicating bits between the substrate-level fixes landed this cycle and behavioural-runtime evidence on the goal-pipeline / survival-leg / cue-authority axes. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding.

Immediate Work Queue (This Cycle, 2026-06-09)

1. V3-EXQ-603i /failure-autopsy adjudication – pending_review reads 1 item with self-route label substrate_not_ready_requeue flagged precondition_unmet. Adjudicate via /failure-autopsy BEFORE the label drives any governance action; clearing the run for review does not clear the adjudication flag (the manifest’s interpretation is the source of truth – a re-queued successor supersedes it). Output routes to: substrate revisit on the SD-059/MECH-358 affordance-bridge half (if the readiness precondition genuinely held and the bridge’s directed-escape contribution did not lift G_H), or held-pending-substrate (if the precondition failed – the V3-EXQ-653 E2 linker readiness microdiagnostic is the next adjudicating bit).
2. V3-EXQ-653 outcome adjudication – E2 escape-affordance linker readiness microdiagnostic; claim-free, forced-choice 4-action probe over the linker driven directly with controlled E2 consequence proxies + known outcomes; 4 arms x 3 seeds; readiness gates G0-G8. PASS routes back to queue a 603-lineage full behavioural bridge re-test on the linker substrate. FAIL routes to /failure-autopsy on this readiness diagnostic. Does NOT validate/weaken SD-059/MECH-358 or MECH-302/303/304.
3. V3-EXQ-640b outcome adjudication – DLAPTOP-4.local claimed 2026-06-08T06:46Z; the CLEAN EVIDENCE retest of the V3-EXQ-640a cue-authority gain sweep on the post-643a modulatory-authority substrate. PRIMARY acceptance = post_cue_approach_lift > 0 on >=2/3 seeds in the decisive cell ARM_CUE_g5_k10; PASS promotes the diagnostic-only 640a result to EVIDENCE supporting SD-057 / MECH-346 / MECH-347 on the GAP-7 cue-authority axis.
4. V3-EXQ-648a + V3-EXQ-649 substrate-readiness adjudication – the MECH-314a Phase-2 curiosity-channel re-sourcing + the ARC-065 GAP-A shared-channel re-sourcing validation diagnostics; both gate on cand_world_pairwise_dist non-vacuity preconditions. PASS unblocks the MECH-341 committed-class diversity re-test (within-class REPRESENTATIVE diversity readout, NOT committed-class entropy) + the section-8 MECH-314a / MECH-314 / ARC-065 governance updates that stay GATED on PASS.
5. MECH-353 / MECH-354 / MECH-355 follow-ups GATED (carried from 2026-06-06): V3 experiment for MECH-353 (smallest blocked-action env), gated /queue-experiment plan for MECH-354 (cue-authority 640b-routes-the-gating-multipliers), V4 substrate design for MECH-355 (deferred V4-social) – all carried; cue-authority lever still in the 640b PRIMARY-acceptance gate before MECH-354 can build.

Status Snapshot (2026-06-06T01:10Z – nightly docs sync; SD/MECH/ARC landings since the 2026-06-05T19:38Z spec sync: SD-016 Path 3 feedforward cue->slot tagger landed (ree-v3 main 88695ed) – replaces ONLY the slot-SELECTION scores in E1DeepPredictor.extract_cue_context (the saddle-stuck q.k attention V3-EXQ-418i diagnosed at the uniform ln(num_slots) saddle) with a fresh feedforward MLP Linear -> ReLU -> Linear; slot-CONTENT path + cue_action_proj + cue_terrain_proj UNTOUCHED; no new supervised target invented – random MLP sits OFF the saddle so the existing terrain_loss gradient flows into it from step 0; cached last_cue_slot_weights read-only diagnostic + 5 new contracts + V3-EXQ-418m substrate-readiness diagnostic queued (PRIMARY mean selection entropy < 2.5 vs pinned ln(16)~2.773 with the tagger ON). MECH-353 / MECH-354 / MECH-355 affect-cluster mint (REE_assembly master 0a9dda6b99 claims.yaml + 7d89ffd0ba affect_primitives Extension Register consolidation + 9d47b3a945 blocked_agency row + 045dac6b9d MECH-303/304 reuniens-thalamic-relay + remote/recent enrichment + 7c6a1f0b55 MECH-355 soothing decay-acceleration design pass + ea53570ec9 MECH-354 fatigue stop-recover design pass) – three proto-feeling-stream candidates confirmed distinct by the 2026-06-05 affect-stream lit-pulls: MECH-353 z_block/blocked-agency (V3 candidate, v3_pending; SD-029 comparator over action-outcome channel with external-attribution + capacity gates; consumers assert MECH-320 -> decommit MECH-342 gated by ARC-016 -> withdraw at capacity collapse), MECH-354 fatigue stop-recover (V3 candidate, v3_pending; SD-012 side, SD-048 host; Meyniel two-bound accumulator wiring via MECH-342 release actuator + ARC-078 persistence gate; NOT SD-011), MECH-355 soothing autonomic state-gain modulator (V4-social, substrate_conditional, promote/demote suppressed; DECAY-ACCELERATION on MECH-219 z_harm_a recovery_rate + SD-032e drive_bias leak; multiplicative-on-existing-state so soothing != sedation). All three doc + claims-only – NO substrate code. MECH-303 + MECH-304 enrichment amend (REE_assembly master 045dac6b9d) adds the midline-thalamic relay (nucleus reuniens -> BLA) + remote-vs-recent time-since-encoding dependence surfaced by Silva 2021 lit anchor; option (i) enrichment + named candidate third-sub-mechanism flag instead of minting on a single rodent anchor. goal_pipeline:GAP-7 frontmatter correction (REE_assembly master f798bd1b80) – closure-map node rewritten to reflect that L2-L3-L4 + L6-L7 substrate already landed 2026-06-04 as SD-057 (MECH-344..348); resume_condition repointed at the real remaining work (L9 wanting!=liking acceptance, GAP-2-gated + in-flight 637/640a validation). SD-037 axis (a) + axis (b) closure_plan frontmatter (REE_assembly master 45f70df66d) – both SD-037 plan-of-record docs now render as real Explorer Closure-tab cards instead of frontmatter_pending placeholders. Thought-intake sweep cleared to 0 unprocessed (REE_assembly master 6326680a24 + 56f97b5184 + d962a67c5b + 7e8d785412) – 8 remaining sweep-unprocessed PARTIALs handled (4 mark-only incl. action_object_type_abstraction + action_policy_and_multi_goal + ocd4 + plasticity_window_neuromodulators; 4 forward-content intakes: orienting_surveying_drive, therapy_action_pressure, cross_version_missing_bits V4/V5, grammar_llms_v5_primitive_mining); 12 incorporated-but-unmarked thoughts backfilled. V3-EXQ-641 coherence-ablation experiment built + queued (ree-v3 main aaa0bb1) – single harness-level diagnostic settling BOTH 2026-04-23 binding + path_integral intakes on the shared discriminator (is C(tau) non-reducible to E(tau)?); paired A/B selectors over identical hippocampal-rollout candidate pools (A = pure error-minimiser argmin E; B = coherence-weighted argmin(E + lam*(-log C)) with GAP-RELATIVE coherence authority gain=0.5 so coherence is competitive in near-ties but subdominant when E-gap exceeds gain*range); pre-flight probe confirmed the fix produces 44/80 flips; acceptance D1 frac_state_div>=0.05 AND D3 |Spearman corr(E,C)|<0.90 AND SPEC real_div>=rand_div+0.05; 3 conditions (real_C_clean / rand_C_control / real_C_perturb) x 3 seeds. V3-EXQ-640a SD-057 cue-AUTHORITY GAIN SWEEP queued (ree-v3 main e4a25e5) – 2-axis factorial cue_recall_gain x incentive_drive_kappa_weight on scaffolded_sd054_onboarding; gates the planned V3-EXQ-638b interoceptive need-gating substrate. V3-EXQ-610f INV-074 / MECH-333 / MECH-334 crystallization-necessity TRUE-NEGATIVE-CONTROL retest queued (ree-v3 main 186f310) – supersedes V3-EXQ-610e confounded control (610e D2 control did NOT collapse because MECH-313 noise floor + MECH-341 E3 score-diversity + always-on policy entropy bonus held diversity up in BOTH arms); ARM_0 strips noise_floor + E3 score-diversity + dACC anti-recency + decouples structured curiosity from crystallize, phase-dependent entropy bonus 0.02 in phases 0-2 vs sweep {0, 0.005, 0.02} in phase 3, ARM_4 floor-on for MECH-341/313 contrast; 5 conditions x 3 seeds, ~35h cloud run; 10th crystallization-arm result with first genuine true-negative control. evidence/experiments/ flat top-level holds 285 v3_exq* manifests; per-machine fleet runner_status/ DLAPTOP-4.local 610 + Daniel-PC 28 + EWIN-PC 77 + cloud-1..4 254/199/154/150 + ree-worker-3 133 completed queue_ids; pending_review.md 0 items at this read; queue items[] 3 items all claimed (V3-EXQ-640a ree-cloud-2 / V3-EXQ-610f ree-cloud-1 / V3-EXQ-641 DLAPTOP-4.local))

• Substrate / governance landings since the 2026-06-05T19:38Z spec sync:
  • SD-016 Path 3 (feedforward cue->slot tagger) (e1.cue_slot_tagger, ree-v3 main 88695ed) – addresses the V3-EXQ-418i diagnosis that Path 1 (auxiliary diversification loss) was “insufficient regardless of weight; the attention bottleneck is categorically in query selectivity, not slot orthogonality.” Replaces ONLY the slot-SELECTION scores in E1DeepPredictor.extract_cue_context; slot-CONTENT path (value_proj -> output_proj -> cue_context) + cue_action_proj (449a z_world concat retained) + cue_terrain_proj UNTOUCHED. New cue_slot_tagger = Linear(world_dim, hidden) -> ReLU -> Linear(hidden, num_slots) MLP producing non-uniform logits from step 0 so it sits OFF the uniform ln(num_slots) saddle that the legacy q.k attention was pinned at – the existing terrain_loss gradient now flows back into it and shapes contextual selectivity. No new supervised target invented; same gradient source. Read-only diagnostic _last_cue_slot_weights cached on E1DeepPredictor for validation experiments. Config (E1Config + REEConfig.from_dims; all no-op defaults): sd016_cue_slot_tagger (False; requires sd016_enabled=True), sd016_cue_slot_tagger_hidden (32), sd016_cue_slot_tagger_temperature (1.0). Off path bit-identical (selection entropy == ln(16) exactly). Honest scope: restores RETRIEVAL selectivity; full action_bias_div >= 0.05 behavioural propagation also depends on cue_action_proj (separate SD-055 differentiable-CEM / ARC-065 concern). MECH-094 N/A (waking E1 query). 7/7 preflight + 5/5 new contracts PASS. V3-EXQ-418m substrate-readiness diagnostic queued (PRIMARY acceptance = mean selection entropy < 2.5 vs the pinned ln(16) ~ 2.773 with the tagger ON; SECONDARY = cue_context per-channel std + safe-vs-dangerous action_bias_div).
  • MECH-353 (blocked_agency / z_block) (REE_assembly master 0a9dda6b99 + 9d47b3a945 + 7d89ffd0ba; ARC-016 / MECH-320 / MECH-342 / SD-029 / MECH-112 / SD-011 / SD-019b deps) – V3 candidate / v3_pending, third proto-feeling stream confirmed distinct (not reducible to harm SD-011 / suffering SD-019b / residue MECH-056 / commitment-block MECH-090) by the 2026-06-05 boundary-violation / blocked-agency / anger lit-pull (5 anchors: Davis & Montag 2019 RAGE primary-process distinct, Papini 2024 FNR antecedent zero noxious input, Bertsch 2020 prefrontal-analogue, Steindl 2015 reactance != helplessness, Carruthers 2012 sense-of-agency comparator). Detector = SD-029 comparator on the action-outcome channel; antecedent = expected-minus-realised outcome (FNR; MECH-112 expectation); smallest computational form = integrated comparator-mismatch + external-attribution gate + capacity gate. Sharp dissociation from Q-036 suffering pole: capacity-COLLAPSED -> withdraw (z_harm_a; already encoded) vs capacity-RETAINED -> ASSERT (the missing piece). Consumers: assert / escalate-effort (raise MECH-320 vigor) -> decommit (MECH-342) gated by ARC-016 if assertion fails -> hand off to suffering accumulator (SD-019b / z_harm_a) at capacity collapse. The V4 social/coercion-pole Stream B (coercion / domination / injustice) deferred as V4 stub (needs other-agent model; Steindl 2015 reactance + injustice appraisal). V3 discriminative experiment proposed: smallest blocked-action env (env repeatedly blocks an intended predicted-to-succeed action with harm + goal held constant, measure z_block rise + assert/persist distinct from withdraw, dissociation from z_harm_a under matched controllability). NO substrate code; NO experimental evidence yet.
  • MECH-354 (effort / fatigue stop-recover two-bound accumulator) (REE_assembly master 0a9dda6b99 + ea53570ec9 design pass; SD-012 / SD-048 / MECH-342 / ARC-078 / SD-017 / SD-011 deps; NOT SD-011) – V3 candidate / v3_pending, two-bound (hysteretic) leaky cost-evidence accumulator (Meyniel 2013) on the SD-012 homeostatic-drive side, substrate-hosted by SD-048 interoceptive channel. F += Se*effort during exertion, F -= Sr during rest; STOP at upper bound, recover to lower bound before re-engaging; Se / Sr / bound-gap incentive-modulated. Two time-constants: fast within-task accumulator + slow Process-S sleep-pressure variant whose recover phase is OFFLINE (SD-017). Wiring: in-task STOP becomes a NEW deficit input to the MECH-342 release actuator (deficit_f OR-composed with the R-c execution-readiness deficit) + feeds ARC-078 goal-disengagement via a cost / benefit (non-aversive) channel; recover / re-accumulation half lives on SD-012 / SD-017, NOT MECH-342. Decisively NOT an SD-011 / z_harm stream (must stay off the SD-048 fatigue-drift -> harm_obs_a additive path). GATED /queue-experiment plan recorded (acceptance: time-on-task disengage independent of harm / controllability, incentive-reversible, WITH a rest-recovery phase). CONTAINMENT GATE honoured: build + queue gated behind cue-authority / z_goal work (V3-EXQ-640a review -> 638b route) because the incentive coefficients multiply the wanting signal 640a is calibrating. Design + gated plan ONLY; no ree_core / claims.yaml / experiment-queue edits beyond the candidate registration.
  • MECH-355 (soothing autonomic state-gain modulator) (REE_assembly master 0a9dda6b99 + 7c6a1f0b55 design pass; MECH-219 / SD-012 / SD-032e / SD-011 deps; NOT 302/303/304/112) – V4-social, candidate, substrate_conditional (promote / demote suppressed). Resolved both verdict-unsettled questions from the 2026-06-05 relief / safety / soothing lit-pull. Q1 UPDATE RULE: DECAY-ACCELERATION primary (recovery-side multiplier on MECH-219 z_harm_a recovery_rate + SD-032e drive_bias leak; onset / sensory streams untouched) + optional default-OFF / ablatable gain-reduction secondary face on the SD-032e accumulation WRITE only (Heinrichs response-suppression). Rejected setpoint-shift (sedation failure mode) and gain-reduction-as-primary (prediction-adjacent -> safety collapse). Key safety property: multiplicative-on-existing-state -> zero effect on a calm agent -> soothing != sedation. Substrate sharpening: route via SD-032e drive_bias (affective component) NOT SD-012 base drive_level / z_harm_s / z_harm_un. Q2 SCOPE: defer soothing substrate to V4-social (every V3 endogenous trigger redundant-with-recovery_rate or collapses-into-safety); pre-register the trigger-pluggable mechanism. SEPARATE V3-tractable “endogenous parasympathetic recovery / autonomic rebound” sibling claim surfaced (offset-triggered; same update rule; endogenous driver) to close the sympathetic-rich / parasympathetic-poor asymmetry – gated on user decision + own lit-pull, NOT folded into MECH-355. Dissociations from relief (302) / safety (303/304) / wanting (112) preserved; falsification signatures listed.
  • MECH-303 / MECH-304 reuniens-thalamic-relay + remote / recent enrichment (REE_assembly master 045dac6b9d) – amends the safety cluster with the midline-thalamic relay (nucleus reuniens -> BLA) + remote-vs-recent (time-since-encoding) dependence surfaced by the Silva 2021 lit anchor (Nat Neurosci, DOI 10.1038/s41593-021-00856-y, conf 0.80). Option (i) AMEND chosen over option (ii) register-third-sub-mechanism: the reuniens relay is anatomically an expression / transmission node fitting naturally as enrichment of the safety-expression pathway (MECH-303 IL / MECH-304 IL->CeA); Silva’s summary lands it most directly on MECH-304 (active cue-specific pole acting ON the fear circuit). The remote-vs-recent dependence IS a genuinely new axis orthogonal to context (303) vs cue (304) – indexes how consolidated the suppressed fear is – but a single rodent anchor off the V3 critical path with no consolidation-interface design does NOT justify a thin third claim (guards over-split per feedback_biology_before_formal_definitions); recorded as a NAMED candidate-third-sub-mechanism flag instead. Edits: claims.yaml MECH-304 functional_restatement + notes + structured enrichment_anchors + revised_utc; MECH-303 cluster-level note safety-expression is prefrontal-hippocampal-THALAMIC; mech_303_contextual_safety_terrain.md + sd_051_conditioned_safety_store.md arch docs; affect_primitives.md safety subsection flag should-consider -> APPLIED + lit-pull status table. Thalamic relay + remote / recent UNIMPLEMENTED (doc-only). build_claims_json.py validate_claims OK (72 invariants); 680 claims (unchanged -> confirms option (i)).
  • goal_pipeline:GAP-7 frontmatter correction (closure_plan node, REE_assembly master f798bd1b80) – the resume_condition + proposed_claims frontmatter still asserted the L2-L3 object-bound incentive-salience substrate was unbuilt and that MECH-BIND-obj / INCENT-token / GOALPTR / CUEWANT / CONSUME were unregistered placeholders. False: landed 2026-06-04 as SD-057 (design doc IMPLEMENTED) with MECH-344 (L2) / 345 (L3) / 346 (L4) / 347 (L6) / 348 (L7) all registered candidate in claims.yaml, implemented in goal.py / agent.py / dacc.py / e3_selector.py, validated by V3-EXQ-636 PASS + V3-EXQ-637 PASS. A new task was mis-scoped off this stale field to re-build SD-057 as a duplicate. Updated proposed_claims -> registered IDs, rewrote resume_condition to point at the real remaining work (L9 wanting!=liking acceptance, GAP-2-gated + in-flight 637 / 640a validation), added l2l3_l6_l7_landed_2026_06_04 provenance field. Plan-doc frontmatter ONLY – no claims.yaml / substrate / experiment / manifest edits.
  • SD-037 axis (a) + axis (b) closure_plan frontmatter (REE_assembly master 45f70df66d) – both SD-037 plan-of-record docs gain real closure_plan: frontmatter so the Explorer Closure tab renders them as real cards instead of frontmatter_pending placeholders. Axis (a) sd_037_axis_a: 4 nodes (P1/P2 done = empirically unmeetable on fishtank baseline per V3-EXQ-620 zero-distribution + inert phase-2 override block; P3/P4 deferred, cross-linked to axis (b)) -> progress 100% concluded-negatively. Axis (b) sd_037_axis_b: 4 nodes (P1b blocked_pending_substrate owner V3-EXQ-625c [advanced from 625b], cross_plan_link behavioral_diversity_isolation:GAP-A; P2/P3/P4 blocked). Captured the confirmed failure_autopsy_V3-EXQ-625b_2026-06-02 re-route that post-dates the plan bodies: axis (b) env curriculum DID lift z_harm_a (seed 7), C3 failure relocated to behavioural-diversity substrate ceiling (monostrategy lock), routed to /implement-substrate AMEND on that cluster, SD-037 status unchanged.
  • Thought-intake sweep 8 PARTIALs cleared -> 0 unprocessed (REE_assembly master 6326680a24 + 56f97b5184 + d962a67c5b + 7e8d785412) – per-file judgment on 8 remaining sweep-unprocessed PARTIALs. MARKED processed (already incorporated): action_object_type_abstraction + action_policy_and_multi_goal (5+ claims.yaml location: cites each), ocd4 (SD-046 multi-goal candidate), plasticity_window_neuromodulators (canonical-in-place, cross-ref’d by plasticity intakes). INTAKED (forward content not captured): orienting_surveying_drive (pre-approach active-sensing MECH gated behind 638a / 640), therapy_action_pressure (therapy-as-symbolic-commit bridge; rest duplicates ProtoFeelings), cross_version_missing_bits + grammar_llms_v5_primitive_mining (V4/V5 design compasses sharing the event-arc spine). Sweep processed 104 / unprocessed 0 (was 74/30 at session start). INTAKE ONLY – no claims.yaml registration.

  • V3-EXQ-641 coherence-ablation built + queued (ree-v3 main aaa0bb1) – chip task_34494eac. Single harness-level diagnostic settles BOTH 2026-04-23 intakes (binding + path_integral) on their shared discriminator: is C(tau) non-reducible to E(tau)? Paired A/B selectors over identical hippocampal-rollout candidate pools (lockstep via per-step torch reseed before generate_trajectories): A = pure error-minimiser argmin E (substrate’s own softmax(-E/T) rule); B = coherence-weighted argmin(E + lam*(-log C)). E = e3.score_trajectory per candidate; C = in-harness cross-system temporal / phase consistency over the rollout window from Trajectory.world_states (z_world) / .states (z_self) / .action_objects: geom-mean of per-stream V_s temporal-smoothness (MECH-269 form) + z_world <-> z_self delta-alignment (binding term). ree_core/ READ-only. CODE-REVIEW CAUGHT + FIXED THE LOAD-BEARING DEFECT: pre-flight probe found E spread ~O(50) vs -log C spread ~O(0.004) – fixed-lambda C never changes argmin (‘drowning’, the modulatory-bias-selection-authority failure mode). FIXED via GAP-RELATIVE authority (coherence-term range == gain * range(E), gain=0.5; substrate’s modulatory_authority_gain pattern) -> re-probe 44 flips / 80 steps. Added RANGE-MATCHED RANDOM-C CONTROL (coherence-specificity vs generic tie-break) + instantaneous paired-agent world_state divergence metric (behavioural). 3 conditions (real_C_clean / rand_C_control / real_C_perturb) x 3 seeds. diagnostic, claim_ids=[] (bears_on INV-002 / ARC-018 / MECH-061 / MECH-269 / MECH-270 in docstring + manifest only). ACCEPTANCE real_C_clean / seed: D1 frac_state_div>=0.05 AND D3

    Spearman corr(E, C)

    < 0.90 AND SPEC real_div>=rand_div+0.05; PASS >=2/3. GRID: PASS -> register entities/selection.coherence_nonreducibility + /implement-substrate wire C into E3.select; D1+D3-not-SPEC -> specificity_unproven; NOT D1 -> F1 close both; NOT D3 -> F2 close both. Phased: P0=40 warmup (agent.eval, online E2; NO new head) then P1=25 measure; arms split only at P1.

  • V3-EXQ-640a SD-057 cue-AUTHORITY GAIN SWEEP (ree-v3 main e4a25e5; claim_ids=[], diagnostic) – measurement successor to V3-EXQ-640. Routed by failure_autopsy_V3-EXQ-640_2026-06-05: 640 measured the post-cue z_goal / approach trace at a SINGLE operating point (cue_recall_gain=0.2) and settled cue-to-action AUTHORITY missing (cue fires 1557x but moves z_goal ~0.4%, post-cue approach == background; displacement REFUTED). Mechanistic root: cue_recall_gain 0.2 * weak token ~0.2 -> sub-threshold cue_pull ~3 orders too small. 2-axis factorial cue_recall_gain {0.2, 1.0, 5.0} x incentive_drive_kappa_weight {2.0, 10.0} on scaffolded_sd054_onboarding (same ARM_OFF / ARM_CUE_ON family); 7 conditions (1 ARM_OFF + 6 cue-on cells) x 3 seeds. GATES V3-EXQ-638b – do NOT build the 638b interoceptive need-gating substrate until this sweep routes.
  • V3-EXQ-610f INV-074 / MECH-333 / MECH-334 true-negative-control retest (ree-v3 main 186f310; supersedes V3-EXQ-610e) – 610e was the FIRST harness-fixed crystallization run (policy genuinely trained, expansion stepped 1249 params, ewc_penalty in loss) but d2_control_shows_collapse=FALSE because MECH-313 noise floor + MECH-341 E3 score-diversity + the always-on policy entropy bonus held diversity up in BOTH arms -> INV-074’s collapse-without-crystallization premise was never instantiated. 610f makes ARM_0 a TRUE NEGATIVE: (a) strip noise_floor + E3 score-diversity + dACC anti-recency suppression, and decouple structured curiosity from crystallize; (b) phase-dependent policy entropy bonus – ENTROPY_BONUS_TRAIN=0.02 in phases 0-2 (every arm; generates the diversity so D3 sanity is non-vacuous + crystallization has something real to preserve) and SWEPT in phase 3 {ARM_0=0.0, ARM_2=0.005, ARM_3=0.02} to map where the control starts to collapse; (c) ARM_4 = stripped control + noise_floor + E3 score-diversity ON to test MECH-341 / 313’s robust-floor reading honestly (the fork-c observation 610e could only note, now scored via D4). MANDATORY startup assertion (_assert_fixes_wired) ERRORs the run if policy-train / expansion-step / ewc-penalty wiring is missing. Acceptance: D1 ARM_1.end_p3 - ARM_0.end_p3 >= +0.10; D2 ARM_0.end_p2 - ARM_0.end_p3 >= +0.10 (true-negative collapses); D3 both > 0.4 at phase2; PASS = D1 AND D2 AND D3. D4 ARM_4.end_p3 - ARM_0.end_p3 >= +0.10 scores MECH-341 / 313 (per-claim). evidence_direction_per_claim emitted (5 claims). Pre-registered fork: D2 FAIL with D3 holding = V3 policy not WTA-prone -> non_contributory-for-INV-074 (universal invariant NOT weakened), escalate Phase-3 pressure; D2 PASS + D1 FAIL = weakens MECH-334 -> /diagnose-errors. ~35h cloud run (610e actual ~142 min / run x 15 runs). 10th crystallization-arm result; FIRST with both harness fixed AND a genuine true-negative control.
• Immediate Work Queue (This Cycle, 2026-06-06)
  • V3-EXQ-640a SD-057 cue-AUTHORITY GAIN SWEEP – in flight on ree-cloud-2; review applies the 638a-autopsy DISCRIMINATOR GRID at REVIEW to route 638b vs E3 goal_proximity/MECH-295 link revisit vs cue_pull substrate revisit.
  • V3-EXQ-610f INV-074 / MECH-333 / MECH-334 true-negative-control crystallization-necessity retest – in flight on ree-cloud-1; the adjudicating bit between the INV-074 plasticity-injection closure prediction vs the prior 610c/d/e confound-corrupted no-op artefacts.
  • V3-EXQ-641 coherence-ablation – in flight on DLAPTOP-4.local; single experiment settles both 2026-04-23 binding + path_integral intakes on C(tau) non-reducibility.
  • V3-EXQ-418m SD-016 Path 3 substrate-readiness diagnostic – queued via /queue-experiment; primary acceptance mean selection entropy < 2.5 vs pinned ln(16) ~ 2.773.
  • MECH-353 / MECH-354 / MECH-355 follow-ups GATED: V3 experiment for MECH-353 (blocked-action env; smallest discriminative pair) + V3 experiment plan for MECH-354 (gated behind cue-authority 638b / 640a) + V4 substrate design for MECH-355 (deferred V4-social).

Status Snapshot (2026-06-04T01:10Z – nightly docs sync; SD/MECH/ARC landings since the 2026-06-03T01:10Z snapshot: three additional scaffolded_sd054onboarding amends in one day – (1) nursery / feeding scaffold amend (ree-v3 main) adds forced-benefit Stage-0 nursery + survival levers + P2 measurement guard + five-way interpretation grid (routed by failure_autopsy_V3-EXQ-603e-626a-622_2026-06-03 concluding the update_z_goal-wiring amend was necessary-but-insufficient); (2) developmental-window / protected-goal consolidation amend (ree-v3 main) adds Stage-0b protected consolidation window + contact-gated P1/P2 updates so decay_only is reserved for mature tests (routed by V3-EXQ-634 design-error review identifying GoalState.update always-decays z_goal across UNFED steps); (3) seeding-calibration amend (ree-v3 main 15053a3) decouples contact-gating threshold from seeding firing threshold + propagates GoalConfig seeding-magnitude knobs + consumption-event-gated G3 readout (routed by failure_autopsy_V3-EXQ-634b which validated consolidation half but exposed G3 anti-correlated with foraging on seed 42). Modulatory-bias-selection-authority substrate (ree-v3 main) – gap-relative E3.select authority rescales composed modulatory bias to range(mod) == gain * raw_score_range with gain=0.5; primary harm/goal scores NOT modified; necessary-but-not-sufficient for the curiosity lever (604a had curiosity_bias=0.0 genuinely). V3-EXQ-635 substrate-readiness PASS (WITHIN_CLASS lever lift +0.446). MECH-306 promoted candidate -> provisional (REE_assembly master 11c043ea79) – V3-EXQ-627 sustained_drive_trace_validation PASS satisfies the v3_pending gate. Governance evening cycle (REE_assembly master 8c85f06e5a) closed 6 reviews + applied 4 user-approved failure-autopsy dispositions (514l / 632 / 634 / 610c all routed to non_contributory + epistemic_category substrate_ceiling + pending_retest_after_substrate). commitment_closure:GAP-8 SD-033b behavioural validation – V3-EXQ-485b devaluation sensitivity + V3-EXQ-485c task-role discrimination authored + queued (ree-v3 main 9f45b0f); GAP-8 status node blocked -> in-progress. goal_pipeline:GAP-7 incentive-salience ratified into plan-of-record (REE_assembly master db72095d46) – L0-L9 closure map embedded; L1 forced-seed positive-control V3-EXQ-626b queued (ree-v3 main ab55916). commitment_closure:GAP-4 OCD behavioural *b cohort – 7 scripts authored, smoke-tested, and queued at priority 290 against the SD-033b / SD-034 / MECH-266 / MECH-268 / MECH-090 substrates. Closure-drift checker enhancements (REE_assembly master 3133d10723) – lineage-advanced + claims-reclassified-since signals so GAP-2-class stale-since gaps surface. Brain-map visualization multi-pass rebuild – coronal MRI backdrop, region re-drape onto anatomy, three-plane linked view (sagittal + coronal + axial). evidence/experiments/ holds 994 v3_exq* manifests recursive (289 flat top-level); fleet aggregate runner_status/*.json reports 779 unique V3 queue_ids completed across all 8 workers (242 PASS / 406 FAIL / 89 ERROR / 41 UNKNOWN / 1 INCONCLUSIVE; 1598 total before dedup); pending_review.md 0 items at this read; queue items[] 7 items (V3-EXQ-634c seeding-calibration readiness claimed DLAPTOP-4; V3-EXQ-610e INV-074/MECH-333/MECH-334 crystallization-necessity claimed ree-cloud-1; V3-EXQ-463b/464b/466b/467b/468b GAP-4 OCD cohort))

• Substrate / governance landings since the 2026-06-03T01:10Z snapshot:
  • scaffolded_sd054_onboarding amend (nursery / feeding scaffold) (curriculum.scaffolded_sd054_onboarding.nursery_feeding_scaffold, ree-v3 main) – second amend in the scaffolded onboarding chain; routed by failure_autopsy_V3-EXQ-603e-626a-622_2026-06-03 concluding the update_z_goal-wiring amend (2026-06-02) is necessary-but-insufficient: V3-EXQ-603e showed z_goal=0 ecologically across 15 cells because 2/3 seeds never reach foraging competence + the hard P2 env (hazard_food_attraction=0.7) starves benefit_exposure even for survivors; 626a P0 positive control formed z_goal on only 1/3 seeds. Five additive levers (all default no-op; bit-identical OFF): forced-benefit Stage-0 nursery run_stage0_nursery (forced supra-threshold benefit + drive into update_z_goal in dense hazard-free reef-refuge env; positive control “goal stream lights when fed” decoupled from survival skill); scaffold_p1_anneal_hold_fraction lever for staged withdrawal of assistance; explicit STAGE_PLAN module + stage_plan() helper; P2 measurement guard scaffold_p2_hazard_food_attraction_guard override + contact-rate readout distinguishing “infant never fed” from “goal-formation failure despite contact”; substrate-gate + five-way interpretation-branch helpers (substrate_not_engaged / fed_but_no_goal / goal_formed_diversity_inert / goal_formed_mechanisms_load_bearing / goal_formed_behaviour_random_harmful). 731 contracts (19 prior scaffolded + 12 new C6) + 7/7 preflight PASS with master OFF.
  • scaffolded_sd054_onboarding amend (developmental-window / protected-goal consolidation) (curriculum.scaffolded_sd054_onboarding.developmental_window, ree-v3 main) – third amend; routed by the V3-EXQ-634 design-error review. Root cause (verified in code): GoalState.update (ree_core/goal.py:173) ALWAYS decays the persistent z_goal attractor (z_goal *= 1-decay_goal) BEFORE the benefit-gated pull AND REEAgent.reset() does NOT reset goal_state (no goal_state.reset() call in agent.py -> z_goal persists across episodes/phases); the prior scaffold called update_z_goal every step (incl. UNFED steps) in P1/P2 so each unfed step was a pure decay-only washout (x0.995 at decay_goal=0.005). Three additive levers (all default no-op; bit-identical OFF): Stage-0b protected consolidation window (new run_stage0b_consolidation; E1/E2 training open but update_z_goal NOT called so the z_goal attractor cannot be washed out by decay-only updating; retention_gate >= 0.75 of Stage-0 baseline); contact-gated P1/P2 updates (when scaffold_contact_gated_goal_updates is set, _train_episode and _eval_episode call update_z_goal only on validated contact steps; decay_only reserved for mature tests); goal-write-mode constants + per-phase diagnostics (n_contact_refresh_updates, n_decay_only_updates, n_skipped_protected_updates). New C7 contract group; 739/739 contracts + 7/7 preflight PASS with master OFF; bit-identical legacy path verified. V3-EXQ-634b corrected nursery readiness queued -> G0b retention 3/3 PASSed (consolidation amend VALIDATED) but exposed seeding-magnitude / threshold mismatch downstream (G3 anti-correlated with foraging on seed 42) -> routed to next amend.
  • scaffolded_sd054_onboarding amend (seeding-calibration + consumption-gated G3) (curriculum.scaffolded_sd054_onboarding.seeding_calibration, ree-v3 main 15053a3) – fourth amend in one day; routed by failure_autopsy_V3-EXQ-634b. Consolidation half VALIDATED (G0b retention 3/3, n_decay_only_updates=0) but exposed benefit-magnitude / threshold mismatch (verified in code): contact-gating skipped only benefit <= contact_threshold (1e-6) but GoalState.update (goal.py:209-224) seeds only when effective_benefit = benefit * z_goal_seeding_gain(1.0) * (1 + drive_weight(2.0) * drive_trace) > benefit_threshold(0.1); natural wild benefit (obs_body[11] ~0.03) stays sub-threshold, so the band (1e-6, ~0.1-effective) DECAYED the consolidated trace during real foraging instead of refreshing it – the forced 1.0 nursery feed was the only supra-threshold input. 634b seed 43 (475 P2 contact-refresh calls, contact_rate 0.348) collapsed z_goal to ~4.5e-05 while non-foraging seed 42 “passed” G3 by carrying the untouched forced-feed nursery trace (0.4398). Three coupled fixes (all default no-op; bit-identical OFF): decoupled contact-gating threshold (scaffold_contact_gating_benefit_threshold sentinel < 0 -> reuses readout threshold; >= 0 separates the gating floor from the contact-RATE readout so sub-seeding whiffs are PROTECTED not decay-only updated); goal-seeding magnitude propagation (Optional scaffold_z_goal_seeding_gain / scaffold_benefit_threshold / scaffold_drive_floor knobs propagated onto agent.goal_state.config at the top of each seeding-capable stage – GoalConfig owns the magnitudes per MECH-186/187/188 / SD-012 precedent; scaffold propagates so the 634c sweep can vary them through the scaffold’s own config surface); consumption-event-gated G3 readout (P2OnboardingMetrics gains z_goal_norm_at_contact_peak + num_contact_events; G3 reads at genuine seeding events so a z_goal=0-at-contact read is no longer masked by the carried forced-feed nursery trace). New C8 contract group (6 contracts); 744/744 contracts + 7/7 preflight PASS with master OFF + amend OFF; bit-identical legacy path verified. V3-EXQ-634c multi-arm sweep over {z_goal_seeding_gain, benefit_threshold, drive_floor} x strengthened P0/P1 budgets queued (claimed DLAPTOP-4.local 2026-06-03T22:04Z); ready stays false; do NOT queue 603f until 634c clears a consumption-event-gated gate.
  • modulatory-bias-selection-authority substrate (ethics_engine_3.modulatory_bias_selection_authority, ree-v3 main) – gap-relative E3.select authority for modulatory / diversity score-bias channels. Root cause (604a / 624a / 614d cluster autopsy): fixed small modulatory magnitudes (~0.05-0.1) added to primary scores whose raw_score_range was much larger never changed the argmin – 604a curiosity_bias=0.0 every arm, 624a vigor action_density byte-identical ON==OFF, 614d within-class temperature -> committed-class entropy byte-identical across T=0.5/1.0/2.0. Approach (b) gap-relative scaling (user-confirmed AskUserQuestion 2026-06-03): rescale composed modulatory bias so range(mod) == modulatory_authority_gain * raw_score_range; takes precedence over legacy normalize_score_bias_to_e3_range when on. Sibling stratified-across-class normalization in e3_score_diversity.stratified_select normalizes class-representative scores to UNIT range before the stratified_temperature softmax (614d C2 fix – absolute class-rep gap no longer collapses committed-class selection). SAFETY: primary scores NOT modified -> commit-threshold / running_variance / softmax-temperature / urgency-interrupt / MECH-090 admission semantics unchanged; gain=0.5 < 1.0 keeps modulatory competitive in near-ties but subdominant when the primary harm/goal gap exceeds gain*range (clearly-harmful candidate stays rejected). Config (REEConfig + from_dims + E3Config, all default no-op): use_modulatory_selection_authority + modulatory_authority_gain (0.5) + modulatory_authority_min_range_floor (1e-6). NECESSARY-BUT-NOT-SUFFICIENT for the curiosity lever (604a had curiosity_bias=0.0 genuinely zero – MECH-314a no active residue centers + 314b/c broadcast-by-design). 734/734 contracts + 7/7 preflight PASS with flag OFF (regression-clean under two pytest-randomly orderings). V3-EXQ-635 substrate-readiness PASS 2026-06-03 (WITHIN_CLASS lever lift +0.446, harm down, 19 authority-normalized ticks). Unblocks per-claim EVIDENCE retests of MECH-314 / MECH-320 / MECH-341 + the MECH-343 hypothesis. Concurrency note: clearing the substrate gate during the 614d review auto-spawned IGW-024 for this substrate; the two sessions converged on the identical design and the joint working-tree implementation was landed from the interactive session (igw-024 stood down, empty worktree).
  • MECH-306 promoted candidate -> provisional (mesolimbic_dopamine_sustained_drive_trace, REE_assembly master 11c043ea79) – V3-EXQ-627 sustained_drive_trace_validation (experiment_purpose=evidence; PASS; supports; exp_conf=0.773 > 0.62 gate) was ALREADY landed + reviewed earlier in the morning 06:33Z cycle. The prior /governance hold was purely the manual v3_pending gate, whose purpose (await a V3 evidence run) is now satisfied. status candidate_substrate_landed -> provisional, v3_pending removed, promoted_utc 2026-06-03. Promoted to PROVISIONAL not active (genuine_exp_count=1, lit_conf=0, novel_discovery quadrant; full active deferred pending replication). implementation_phase:v3 retained – it no longer gates because the indexer holds on implementation_phase=v3 only when zero V3 runs exist (627 is a V3 run). MECH-306 now out of the pending_user decision queue (0 pending_user).
  • Governance evening cycle (REE_assembly master 8c85f06e5a) – ran /governance evening cycle 2026-06-03 end-to-end (747 files). 6 pending walked (2 PASS, 4 FAIL); all 4 FAILs had confirmed same-day failure-autopsies, user-approved (AskUserQuestion) to apply verbatim. 514l (SD-049/SD-015/MECH-229/MECH-230) + 632 (MECH-230): non_contributory + epistemic_category substrate_ceiling + pending_retest_after_substrate (foraging-competence/benefit-contact ceiling; 632 seed42 clean positive z_goal_norm=3.0115 corrects recorded gap to foraging-competence upstream not goal-projection wiring; author-stamped weakens/does_not_support overridden; narrow_supports_flag). 634 (diagnostic, claim_ids=[]): non_contributory + substrate_ceiling (nursery G0 forced-feed PASS 3/3, wean-to-wild G1/G2/G3 FAIL, substrate_not_engaged). 610c (INV-074/MECH-333/MECH-334): non_contributory + measurement_test_design_defect (manifest) + pending_retest (HARNESS defect: policy never trained, crystallize freezes untrained heads, ewc_penalty never in loss -> behavioral no-op; INV-074 universal invariant NOT weakened; substrate_queue action=none; route /queue-experiment 610d -> 610e). PASS: 633 (MECH-094, evidence, supports 3/3) corroborates stable MECH-094; 635 (diagnostic) validates modulatory-bias-selection-authority. closure_drift 0 drifted / 4 suppressed / 0 missing – clean.
  • commitment_closure:GAP-8 SD-033b behavioural validation – GAP-8 verify (audit confirmed V3-EXQ-485b/485c never ran, not silent-drop; case (c) never-claimed; IDs reusable). Blocker GAP-3 (env extensions) had landed 2026-05-17 so GAP-8 was already unblocked. Authored+smoke+queued both as REPRESENTATION-LEVEL MECH-263 functional-signature diagnostics via /queue-experiment (ree-v3 main 9f45b0f): 485b devaluation sensitivity (AVERSIVE z_harm devaluation at fixed z_world; smoke 3/3 PASS) + 485c task-role discrimination (same-z_world / different-task-stage histories; smoke 4/4 PASS). Both direct-drive like 485/485a; experiment_purpose=diagnostic; claim_ids=[SD-033b, MECH-263]. Frozen-zeroed OFC head -> behaviour not measurable; FULL promotion candidate->provisional still needs the deferred trained-head arm. GAP-8 lands PARTIAL on 485b/c PASS. GAP-8 plan node blocked -> in-progress + status-row + EXQ-table + decision-log updates (REE_assembly master 7c1db828a2).
  • goal_pipeline:GAP-7 incentive-salience ratified into plan-of-record (REE_assembly master db72095d46) – user ratified the GAP-7 proposal (thought_intake_2026-06-01_goal_wanting_liking_stream_repair.md Section 9) into the plan-of-record. Added live goal_pipeline:GAP-7 node (object-bound incentive-salience L2-L3 + L1 harness positive control + L7 consumer-readout audit; status open, load-bearing, owner null, depends_on GAP-2) to goal_pipeline_plan.md: frontmatter node + gap-inventory row + Phase 7 section + status-table row + decision-log entry + embedded the L0-L9 closure map table into the plan body (was intake-doc-only). YAML re-parse verified (7 nodes; GAP-7 depends_on goal_pipeline:GAP-2, 5 proposed_claims placeholders). Did NOT register MECH-BIND-obj / MECH-INCENT-token / MECH-GOALPTR / MECH-CUEWANT / MECH-CONSUME in claims.yaml (placeholders only); no substrate_queue entry; no experiment scripts. L1 forced-seed positive-control V3-EXQ-626b queued (ree-v3 main ab55916) – 626a wired update_z_goal (Class-1 fix) + the F0 unit positive control passes 6/6, but 626a’s ARM_A positive control drew benefit from ecological foraging so seeds that never foraged showed z_goal=0; 626b adds a genuine FORCED-SEED positive-control arm (decoupled from foraging via run_stage0_nursery pattern) so the developmental-window diagnostic SEES a non-zero stable z_goal in a positive-control arm depending only on the harness fix, NOT on the GAP-2 foraging-competence substrate.
  • commitment_closure:GAP-4 OCD behavioural *b cohort – 7 scripts authored + smoke-tested + queued at priority 290 against the SD-033b (OFC), SD-034 (closure operator), MECH-266 (asymmetric hysteresis), MECH-268 (dACC conflict saturation), and MECH-090 (commit entry) substrates. Substrate side resolved 2026-06-02 (MECH-090 + MECH-342 validated by 592g); runs on GAP-3 CausalGridWorldV2 env extensions + GAP-11 committed_mode_curriculum. Cohort: V3-EXQ-460b (SD-034 verified-but-not-released), 461b (MECH-090 + SD-033a delayed reward persistence), 463b (MECH-268 dACC conflict saturation), 464b (MECH-266 competing goals), 466b (SD-034 satisficing residue discharge), 467b (MECH-266 mode stickiness), 468b (SD-034 + MECH-268 commitment vs contradiction). 5/7 already claimed by ree-cloud workers at snapshot time; 467b + 468b pending.
  • closure-drift checker enhancements (REE_assembly master 3133d10723) – added two date-aware drift signals to scripts/check_closure_drift.py so the closure-drift check (governance.sh Step 3c) catches the staleness class that hid goal_pipeline:GAP-2: (1) owner_exq lineage-advanced (later-lettered sibling has terminal evidence), (2) unblocks_claims reclassified by a confirmed failure-autopsy dated after the node’s last_updated. Both run for ALL non-terminal nodes incl. suppressed. New “Stale since last update – review” report section; warn-only, not counted as drift. Verified pre-fix GAP-2 (owner 514g / last_updated 2026-05-16) fires BOTH signals; post-fix GAP-2 (owner 514l / 2026-06-03) clean on both. Live run: 4 genuine review nodes surfaced (arc_062 GAP-H, goal_pipeline GAP-4, self_attribution GAP-1, behavioral_diversity_isolation GAP-B), 0 GAP-2 false-positive, low noise.
  • Brain-map visualization rebuild – coronal MRI backdrop replaces the hand-drawn cartoon (511KeV Wikimedia, CC BY-SA 4.0; previously sagittal cartoon); full region re-drape onto coronal anatomy (cortex along mantle, thalamus / BG / cingulate over central gray + ventricles, hippocampus / amygdala over temporal lobes, PAG / respiratory at brainstem, insula in sylvian fissures) with naturally-bilateral regions made symmetric L/R pairs; pre-cropped coronal MRI to brain (removed face / neck); 1:1 gridline-measured alignment fixes (PFC / motor seated on cortex mantle, cingulate above ventricles, thalamus below). Linked 3-plane view (sagittal + coronal + axial MRI) emitted from a single region table via new generator build_brain_planes.py; rewrote brain_map.html to load all three + a substrate column, fill+select a region across every plane it appears in, shared sidebar; verified with jsdom runtime test. Multiple user-feedback iterations across the day.
• Runner activity since the 2026-06-03T01:10Z snapshot: per-machine runner_status/*.json now aggregates to 779 unique V3 queue_ids completed across all 8 workers after dedup by queue_id (242 PASS / 406 FAIL / 89 ERROR / 41 UNKNOWN / 1 INCONCLUSIVE; 1598 total completion records before dedup, reflecting cross-machine retries + smoke runs). DLAPTOP-4.local alone shows 610 completed queue_ids in its per-machine status JSON; evidence/experiments/ contains 994 v3exq* manifests on disk (recursive count incl. nested per-run dirs; 289 flat top-level). First nightly snapshot that has aggregated the per-machine status JSONs as recommended in yesterday’s note.
• Queue depth: 2 -> 7 items at this read – V3-EXQ-634c scaffolded_sd054_onboarding seeding-calibration 4-arm readiness diagnostic (claimed DLAPTOP-4.local 2026-06-03T22:04Z; supersedes 634b); V3-EXQ-610e INV-074 / MECH-333 / MECH-334 crystallization-necessity retest with three-prescription harness fix (claimed ree-cloud-1 2026-06-03T21:19Z; supersedes 610d); V3-EXQ-463b MECH-268 dACC conflict saturation (claimed ree-cloud-4); V3-EXQ-464b MECH-266 competing goals (claimed ree-cloud-2); V3-EXQ-466b SD-034 satisficing residue discharge (claimed ree-cloud-3); V3-EXQ-467b MECH-266 mode stickiness (pending); V3-EXQ-468b SD-034 + MECH-268 commitment vs contradiction (pending). Five of the seven are the GAP-4 OCD behavioural *b cohort queued late in the session against the GAP-3-landed CausalGridWorldV2 env extensions + GAP-11 committed_mode_curriculum.
• Pending review: 2 -> 0 items at this read – all walked via two /governance cycles (16:59Z + 19:57Z).
• Bottleneck (continuation from yesterday’s framing): the ecological-evidence v3_pending lift requirement remains the dominant blocker, but specific axes advanced today – (a) the scaffolded_sd054_onboarding substrate has now been amended THREE additional times in one day (nursery -> developmental-window -> seeding-calibration) on top of yesterday’s wiring amend; runtime readiness is converging on a clearable substrate; V3-EXQ-634c is the adjudicating bit; (b) the modulatory-bias-selection-authority substrate now lets MECH-314 / MECH-320 / MECH-341 modulatory levers actually influence the committed argmin via gap-relative scaling, so pending behavioural re-runs on these claims can produce non-vacuous evidence; (c) MECH-306 has cleared the V3-pending gate and is now provisional – one substrate-conditional hold resolved. V3-EXQ-610e (INV-074 crystallization-necessity) remains the adjudicating bit between the INV-074 plasticity-injection closure prediction vs the prior 610c/d untrained-policy no-op artefact – 610e wires REINFORCE policy training + stepped expansion-parameter optimizer + EWC penalty into the Phase-3 loss, each verified by a mandatory startup assertion. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding.

Immediate Work Queue (This Cycle, 2026-06-04)

1. V3-EXQ-634c outcome adjudication – DLAPTOP-4.local claimed 2026-06-03T22:04Z; multi-arm sweep over {z_goal_seeding_gain, benefit_threshold, drive_floor} x strengthened P0/P1 budgets on the seeding-calibration amend. PASS clears the runtime-readiness gate for scaffolded_sd054_onboarding and unblocks the V3-EXQ-603f re-issue; FAIL routes to substrate revisit on the foraging-competence half (G1 P1-survival was the seed-43 gap last cycle).
2. V3-EXQ-610e outcome watch – ree-cloud-1 claimed 2026-06-03T21:19Z; supersedes 610d untrained-policy no-op signature with three prescriptions each verified by mandatory startup assertion (real REINFORCE on gated_policy as behavioral policy in Phases 0-2; stepped expansion_parameters() optimizer post-crystallize in Phase 3; residue_field.ewc_penalty() summed into the Phase-3 loss after snapshot_ewc_anchor). 2-arm / 3-seed / IGW-023 Phase-3 pressure / D1/D2/D3 acceptance unchanged from 610d. Pre-registered substrate-ceiling fork (docstring grid row c): trained non-uniform policy + ARM_0-no-collapse STRENGTHENS MECH-341 / MECH-313, does NOT weaken INV-074.
3. GAP-4 OCD cohort outcome landing watch – V3-EXQ-463b / 464b / 466b / 467b / 468b ecological behavioural runs on the substrate-validated MECH-090 + MECH-342 + MECH-266 + MECH-268 + SD-034 stack. These form the next governance batch for ecological-evidence v3_pending lift on the OCD-axis claims.
4. V3-EXQ-485b / 485c outcome watch (commitment_closure:GAP-8) – SD-033b behavioural validation via MECH-263 functional-signature diagnostics; frozen-zeroed OFC head + direct-drive limits to representation-level (not behavioral) for now. Trained-head arm deferred (parallel SD-033a GAP-1) – PARTIAL on 485b/c PASS.
5. V3-EXQ-626b outcome watch (goal_pipeline:GAP-7 L1) – forced-seed positive-control on the harness-fix; decoupled from foraging so the developmental-window diagnostic sees a non-zero stable z_goal depending only on the harness fix (not GAP-2 foraging-competence). On PASS, L7 consumer-readout wiring audit (dACC does NOT read z_goal directly) is the next GAP-7 deliverable.

Status Snapshot (2026-06-03T01:10Z – nightly docs sync; SD/MECH/ARC landings since the 2026-06-02T01:10Z snapshot: MECH-090 R-c continuation Phase-2 env-source follow-on (ree-v3 main fa026a0 substrate + 60d1a90 V3-EXQ-630 queue + e9e1b2b doc-id-fix; REE_assembly master b23ad1a125 + 6be3673781) closes the named Phase-2 follow-on – env emits info[mech090readiness_outcome] = clip(1 - mean(limb_damage), 0, 1) under env-only kwarg mech090_readiness_outcome_enabled (default False ABSENT-WHEN-DISABLED); REEAgent.sense forwards into commit_readiness.update; agent UNCHANGED consumer+seam (the 2026-05-29 pass) now has a real source. MECH-342 maintenance-time commitment-release coupling B3b (ree-v3 main 780d12f + REE_assembly master 625e218779) – release-side complement to MECH-090 admission predicate, closes V3-EXQ-592f reach gap. Pure-arithmetic regulator commit_maintenance_release.py; OR-composition of decisiveness + nav_competence deficits, drift-to-bound + reengagement leak; distinct from MECH-090 / MECH-091 / ARC-028 / MECH-269b / MECH-340; V3-EXQ-592g PASS validation 2026-06-02 (all 6 criteria). scaffolded_sd054_onboarding AMEND update_z_goal wiring + Stage-0 positive control (ree-v3 main deb24cc + d09af0e; REE_assembly master 36b0130ecf) – root-cause fix for V3-EXQ-603d Class-1 harness/wiring artifact LIVING IN THE SUBSTRATE MODULE; TWO-PART FIX wiring + z_goal_enabled=True+drive_weight=2.0 config. MECH-341 stratified_within_class_temperature amend (ree-v3 2026-06-01) – (a) within-class proportional sampling lever to decouple Layer B within-class sub-axis from across-class sub-axis; (b) A-vs-B partial-redundancy probe naming via existing independent flags. Pre-governance disposition V3-EXQ-592f + V3-EXQ-592g (REE_assembly 01144f9bf6) – 592f re-tagged does_not_support -> non_contributory + epistemic_category substrate_ceiling + cleared pending_retest_after_substrate (reach gap closed by MECH-342); MECH-090 unchanged. Cross-fleet experiment wave queued (ree-v3 main 34e6369 + 829f6b1 + b7fae0a + e9a0b87 + others) – V3-EXQ-627 MECH-306 sustained_drive_trace evidence; V3-EXQ-604a MECH-314 curiosity on SD-056 substrate; V3-EXQ-628 MECH-319 replay/caller_sim falsifier evidence; V3-EXQ-629 MECH-342 ecological; V3-EXQ-630 ARC-029 across-tick. Fleet outage diagnosed + recovered 2026-06-02T23:50Z (hub-writer wedge not migration; all 4 cloud runners crash-looped at startup since ~22:05-23:06Z due to frozen queue snapshot keeping V3-EXQ-610c pending despite FAIL+completed; non-destructive backup+clean recovery; cloud-2/3/4 cut over to PHASE3_COMMANDS_VIA_COORDINATOR=1 same recovery; hub runner intentionally stopped pending operator decision). Plan-doc updates: commitment_closure:GAP-4 owner_exq advanced to V3-EXQ-629 with stale-631 correction (194400a994); arc_062_rule_apprehension:GAP-K owner_exq repointed to V3-EXQ-628 MECH-319 evidence falsifier (6e23af6fc3). Epoch stale-evidence bookkeeping closed (2be6faafd7) – 483c supersession + 17 B.3 nested-manifest stale flags; GOVERNANCE CONSEQUENCE: 483c was sole genuine exp entry for SD-037 / MECH-280 / MECH-281 -> exp_conf for all three -> 0.0; SD-037 retains other active streams (483b / 620b / 625c). Central runner_status decoupled into per-machine files under evidence/experiments/runner_status/ post-2026-05-29 Phase 3 cutover; evidence/experiments/ holds 908 v3_exq* manifests on disk; pending_review.md 2 unclaimed manifests (V3-EXQ-626a + V3-EXQ-625c FAIL non_contributory); queue items[] 2 items (V3-EXQ-603e claimed DLAPTOP-4 + V3-EXQ-614d pending DLAPTOP-4))

• Substrate / governance landings since the 2026-06-02T01:10Z snapshot:
  • MECH-090 R-c continuation Phase-2 env-source follow-on (control_plane.beta_gate.commit_entry_readiness_conjunction.nav_competence.env_source, ree-v3 main fa026a0 + 60d1a90 + e9e1b2b; REE_assembly master b23ad1a125 + 6be3673781) – closes the named Phase-2 follow-on from the 2026-05-29 R-c continuation landing. The 2026-05-29 pass wired the consumer + notify_outcome seam but grep-verified ZERO callers ever pushed via the seam (committed_mode_curriculum.py computes nav_competence but never pushes), so in any ecological run the readiness EMA sat pinned fail-open at 1.0 and the across-tick axis added no signal – exactly why V3-EXQ-063a left it OFF. This pass adds the env source. CausalGridWorldV2 gains env-only kwarg mech090_readiness_outcome_enabled (default False; NOT in REEConfig.from_dims, matches SD-022 / SD-023 / SD-029 / SD-047 / SD-048 / SD-049 / SD-054 precedent) emitting info[mech090_readiness_outcome] = clip(1.0 - mean(limb_damage), 0, 1) – a Cisek-Kalaska affordance-preparation / motor-program-readiness scalar that degrades on SD-022 limb damage and recovers on heal. REEAgent.sense() gains mech090_readiness_outcome arg forwarding the value into commit_readiness.update(). ABSENT-WHEN-DISABLED (no always-present sentinel): default-OFF emits no key, agent reads None, EMA un-advanced. CommitReadiness module UNCHANGED (None-sentinel + simulation_mode gate already supported this). 719/719 contracts + 7/7 preflight PASS; integration smoke (aggressive all-limb scheduled injection magnitude 0.5) drives readiness EMA to 0.001 below floor 0.3 and recovers on heal. V3-EXQ-630 queued ecological 3-arm (OFF / GATED_NAV_COMP_ON / GATED_BOTH_ON) on SD-022 scheduled-injection env; claim_ids=[ARC-029, MECH-090]; C3-FAIL routes to /failure-autopsy NOT /diagnose-errors; 630 was claimed and dropped from the queue snapshot mid-day during the fleet outage recovery.
  • MECH-342 maintenance-time commitment-release coupling B3b (control_plane.commit_maintenance_release, ree-v3 main 780d12f + REE_assembly master 625e218779) – release-side complement to the MECH-090 commit-entry R-c admission predicate (which the V3-EXQ-592f autopsy + MECH-090 release-path audit + motor-cessation lit-pull established is ADMISSION-ONLY by design). Same two R-c readiness signals MECH-090 AND-composes to ADMIT drive a graded bounded-accumulation RELEASE of an already-elevated beta latch when they degrade mid-commitment. Closes V3-EXQ-592f reach gap. Routed by REE_assembly/evidence/planning/mech090_release_path_audit_2026-06-02.md (B1 ruled out: none of ARC-028/MECH-105 completion, MECH-091 urgency, V_s commit-release, SD-034 closure covers degraded-readiness mid-commitment) + targeted_review_mech_090_release_motor_cessation/SYNTHESIS.md verdict B3b. Pure-arithmetic regulator, OR-composition of decisiveness + nav_competence deficits, drift-to-bound + reengagement leak; distinct (falsifiable) from MECH-090 / MECH-091 / ARC-028 / MECH-269b / MECH-340. 700 contracts (685 + 15 new MECH-342) + 7/7 preflight PASS; bit-identical OFF verified. V3-EXQ-592g PASS validation 2026-06-02 (all 6 criteria); V3-EXQ-629 ecological evidence run queued (claimed + dropped mid-day fleet outage). DISPOSITION: MECH-342 stays candidate / v3_pending (592g is diagnostic; v3-pending gate forbids promotion regardless of evidence count; promotion needs an ecological evidence-grade run); substrate_queue MECH-342 status -> implemented_validated_v3_exq_592g.
  • scaffolded_sd054_onboarding AMEND update_z_goal wiring + Stage-0 positive control (curriculum.scaffolded_sd054_onboarding.update_z_goal_wiring, ree-v3 main deb24cc + d09af0e; REE_assembly master 36b0130ecf) – amend (folds V3-EXQ-603d + 625b failure records) on the 2026-05-31 scaffolded_sd054_onboarding substrate. Root cause (confirmed): neither _train_episode nor _eval_episode called agent.update_z_goal -> GoalState.update never reached -> z_goal stayed zero-init every step every arm; V3-EXQ-603d C4 SUBSTRATE_FAILURE was a 626-class Class-1 harness/wiring artifact LIVING IN THE SUBSTRATE MODULE, NOT a substrate ceiling. TWO-PART FIX: wiring + z_goal_enabled=True+drive_weight=2.0 config (603d omitted z_goal_enabled -> goal_state None -> update_z_goal early-returned even with wiring; working reference V3-EXQ-622 sets both explicitly). Stage-0 positive control: two new contracts make a z_goal=0 scheduler structurally unshippable. 19/19 scaffolded contracts + 7/7 preflight + 665/665 full contract suite PASS. V3-EXQ-603e queued (priority 250, supersedes V3-EXQ-603d, EXPERIMENT_PURPOSE=diagnostic, P0/P1=100/50 restored budget, z_goal_enabled=True+drive_weight=2.0); claimed by DLAPTOP-4.local at snapshot time. P0 positive control is the adjudicating bit between harness-bug-fixed and object-binding abstraction-gap hypotheses.
  • MECH-341 stratified_within_class_temperature amend (ethics_engine_3.scoring_trajectory_class_diversity_preservation.within_class_temperature, 2026-06-01) – amend (NOT supersede) of the 2026-05-29 SD-056 t=1 substrate landing routed by failure_autopsy_V3-EXQ-616_2026-05-31 Sections 7 + 10 contingent-on-614b-FAIL-C1 path. Two-part amend: (a) within-class proportional sampling lever via new E3ScoreDiversityConfig.stratified_within_class_temperature (Optional[float], default None = legacy argmin bit-identical) so the A-vs-B probe can dissociate Layer B within-class sub-axis from across-class sub-axis (decoupling avoids un-interpretable single-knob conflation); (b) A-vs-B partial-redundancy probe NAMED via the existing independent master flags use_support_preserving_cem (Layer A) + use_e3_score_diversity (Layer B) which compose to a complete factorial – no new config flag added. When stratified_within_class_temperature is set, within-class sampling is softmax(-class_scores / T) per first-action class before across-class softmax. Three new diagnostics on E3ScoreDiversity.get_state(). 655/655 contracts (645 prior + 10 new amend contracts) + 7/7 preflight PASS bit-identical. V3-EXQ-614c queued -> FAILed instrumentation-defect 2026-06-01T12:45Z -> V3-EXQ-614d corrected-harness re-run queued (pending DLAPTOP-4.local at snapshot).
  • Pre-governance disposition V3-EXQ-592f + V3-EXQ-592g (REE_assembly 01144f9bf6) – 592f re-tagged does_not_support -> non_contributory + epistemic_category=substrate_ceiling + cleared pending_retest_after_substrate (reach gap closed by MECH-342); MECH-090 unchanged (release capability lives on dependent MECH-342); 592g reviewed PASS, MECH-342 validation_note added, substrate_queue MECH-342 status -> implemented_validated_v3_exq_592g. pending_review = 0 indexed after walk.
  • Cross-fleet experiment wave queued (ree-v3 main 34e6369 + 829f6b1 + b7fae0a + e9a0b87 + others) – V3-EXQ-627 MECH-306 sustained_drive_trace 2-arm evidence; V3-EXQ-604a MECH-314 curiosity validation on SD-056 substrate (supersedes 604); V3-EXQ-628 MECH-319 replay / caller_sim falsifier evidence; V3-EXQ-629 MECH-342 ecological; V3-EXQ-630 ARC-029 across-tick. All claimed and dropped from queue snapshot mid-day during fleet outage recovery; results / heartbeats will resurface on origin via the phase3-* writers as workers complete or release.
  • Fleet outage diagnosed + recovered (2026-06-02T23:50Z) – hub-writer wedge (NOT migration), all 4 cloud runners crash-looping at startup since ~22:05-23:06Z due to hub-writer wedge causing frozen queue snapshot to keep V3-EXQ-610c pending despite FAIL+completed status (test_queue_integrity FAIL -> start-limit-hit). Non-destructive recovery via backup+clean tree fix (per reference_hub_writer_wedge.md). Cloud-2/3/4 simultaneously cut over to PHASE3_COMMANDS_VIA_COORDINATOR=1 during the recovery; hub runner intentionally stopped pending operator decision on disable.
  • Plan-doc updates – commitment_closure:GAP-4 owner_exq advanced to V3-EXQ-629 (ecological MECH-342 evidence) with stale-631 correction (REE_assembly master 194400a994); arc_062_rule_apprehension:GAP-K owner_exq repointed to V3-EXQ-628 MECH-319 evidence falsifier (REE_assembly master 6e23af6fc3 – previous owner_exq cited a MECH-341 run by mistake).
  • Epoch stale-evidence bookkeeping closed (REE_assembly master 2be6faafd7) – 483c supersession recorded + 17 B.3 nested-manifest stale flags written. GOVERNANCE CONSEQUENCE: 483c was the sole genuine exp entry for SD-037 / MECH-280 / MECH-281, so exp_conf for all three -> 0.0; SD-037 retains other active streams (483b / 620b / 625c). Scoring-neutral via same-moment diff (only scoring_excluded relabels).
  • Runner / coordinator hardening – runner phantom-PASS fix (ree-v3 8c17c2b) _classify_no_sentinel_result always classifies non-zero-exit + no-sentinel as ERROR (no longer trusts partial stdout verdict: PASS on a crashed process). Phase 3 commands-via-coordinator code landed (ree-v3 main; gates default OFF) replacing the git command-file with POST /commands/issue + /commands/ack; canary on cloud-3 verified end-to-end before the fleet outage; cloud-2 + cloud-4 cut over during the outage recovery; hub still on legacy git command-file pending serve.py PHASE3_COMMANDS_DUAL_WRITE flip + hub PHASE3_COMMANDS_OFF_GIT. serve.py POST /api/machines/<host>/command gained dual-write to coordinator behind PHASE3_COMMANDS_DUAL_WRITE (default OFF).
• Runner activity since the 2026-06-02T01:10Z snapshot: central runner_status.json was decoupled in the 2026-05-29 Phase-3 cutover; per-machine status files under evidence/experiments/runner_status/ are the authoritative surface now (DLAPTOP-4.local + Daniel-PC + EWIN-PC + ree-cloud-1 through ree-cloud-4 + ree-worker-3). Local Mac alone shows 613 completed queue_ids in its per-machine status JSON; evidence/experiments/ contains 908 v3exq* manifests on disk. Cumulative cross-fleet PASS / FAIL / ERROR breakdown not available from a single file at this snapshot – next nightly should aggregate per-machine status JSONs.
• Queue depth: 3 -> 2 items at this read – V3-EXQ-603e Q-045 / MECH-313 / MECH-260 scaffolded_sd054_onboarding 5-arm at restored budget (claimed DLAPTOP-4.local 2026-06-02T06:46Z; supersedes 603d; z_goal_enabled=True + drive_weight=2.0 fix verified pre-run); V3-EXQ-614d MECH-341 within-class temperature corrected-harness 4-arm sweep (pending DLAPTOP-4.local; supersedes 614c instrumentation-defect FAIL). The five ecological evidence runs queued today (V3-EXQ-604a / 627 / 628 / 629 / 630) were claimed and dropped from the queue snapshot mid-day during the fleet outage recovery – they will resurface as results / heartbeats on origin via the phase3-* writers as workers complete or release.
• Pending review: 0 -> 2 unclaimed manifests at this read – V3-EXQ-626a (goal-pipeline developmental-window FAIL) and V3-EXQ-625c (SD-037 axis-b Phase 1b dynamic-crossings FAIL, non_contributory tag). Both are claim_ids=[] diagnostics from concurrent sessions; routing to next /governance cycle.
• Bottleneck (updated framing): the dominant blocker has shifted from the scaffolded_sd054_onboarding harness/wiring gap (now fixed; V3-EXQ-603e in flight) to the ecological-evidence v3_pending lift requirement – multiple substrates (MECH-090 R-c admission + R-c continuation + MECH-342 release-side; MECH-319 simulation-mode rule gate; MECH-306 / MECH-314 score-bias contributors) now have substrate-readiness PASS + diagnostic validation in hand, but the V3-pending governance gate forbids promotion regardless of evidence count until ecological evidence-grade runs PASS. The five ecological evidence runs queued today (V3-EXQ-604a / 627 / 628 / 629 / 630) form the next governance batch. V3-EXQ-603e (Q-045 / MECH-313 / MECH-260 on hook-fixed substrate) remains the adjudicating bit between scaffolded_sd054_onboarding harness-bug-fixed vs deeper z_goal-formation regression; P0 positive control is the adjudicator. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding.

Immediate Work Queue (This Cycle, 2026-06-03)

1. V3-EXQ-603e outcome adjudication – DLAPTOP-4.local claimed 2026-06-02T06:46Z; P0 positive-control gate is the single bit that adjudicates the scaffolded_sd054_onboarding harness-bug-fixed vs deeper z_goal-formation regression. PASS clears Q-045 / MECH-313 / MECH-260 routing and unblocks the V3-EXQ-622 anneal-rate amend; FAIL contradicts V3-EXQ-622 S0 and routes to /failure-autopsy on GoalState.update.
2. V3-EXQ-614d outcome watch – pending DLAPTOP-4.local; corrected-harness re-run of the MECH-341 within-class temperature sweep with committed-class-entropy metric + within-class diagnostics + cross-seed-mean C1 band. PASS routes MECH-341 v3_pending lift; FAIL with branch-active diagnostics routes to substrate revisit.
3. Five ecological evidence runs landing watch – V3-EXQ-604a MECH-314 / V3-EXQ-627 MECH-306 / V3-EXQ-628 MECH-319 / V3-EXQ-629 MECH-342 / V3-EXQ-630 ARC-029 across-tick. All claimed during the day; results will surface via phase3-* writers as workers complete. These form the next governance batch for ecological-evidence v3_pending lift.
4. Hub runner disable decision – the cloud-1 hub runner is intentionally stopped after the fleet outage (it was the wedge culprit). The systemd unit is still enabled. Per OPERATOR_GUIDE, the recommendation is systemctl disable ree-runner on the hub to prevent reboot recurrence. Pending user call on fleet-capacity tradeoff.
5. Hub PHASE3_COMMANDS_OFF_GIT cutover – after serve.py PHASE3_COMMANDS_DUAL_WRITE flip, the hub can move off the git command-file (which is the root cause of its _HEARTBEAT_WRITE=1 requirement and the wedge hazard). Operator-driven; not in this cycle’s autonomous scope.
6. /governance walk of two pending unclaimed manifests – V3-EXQ-626a and V3-EXQ-625c FAIL non_contributory; route to next governance cycle for evidence_direction confirmation + discussed_experiment_dirs mark.
7. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21.

---

Status Snapshot (2026-06-02T01:10Z – nightly docs sync; no fresh SD/MECH/ARC/Q substrate landings; central runner_status freshness ~1h35m at this read (794 cumulative, +8 since 2026-06-01T01:10Z; 195 PASS / 319 FAIL / 87 ERROR / 193 UNKNOWN; deltas PASS +3, FAIL +5, ERROR +0, UNKNOWN +0); two /governance cycles applied 8+2 dispositions and cleared pending_review 1->0 then maintained at 0; queue items[] now 3 / all claimed (V3-EXQ-610c INV-074 / MECH-333 / MECH-334 crystallization-necessity retest on ree-cloud-3, V3-EXQ-626a goal-pipeline harness-fix on ree-cloud-1, V3-EXQ-624 ARC-068 / MECH-320 Niv-vs-Salamone on DLAPTOP-4); five failure-autopsies landed today (V3-EXQ-614c MECH-341 within-class instrumentation defect + V3-EXQ-610a INV-074 / MECH-333 / MECH-334 substrate-ceiling re-read + V3-EXQ-626 Class-1 harness bug + V3-EXQ-603d Class-1 harness/wiring artifact LIVING IN THE SUBSTRATE MODULE + V3-EXQ-592e MECH-090 beta-gate readiness-failure release path missing); /diagnose-errors landed goal_pipeline_tier1.build_config guarded opt-in for SD-011 affective harm stream and queued V3-EXQ-620b/625b corrected runs (620b PASS confirms stream now writes BLA/CeA); substrate_queue gained test_bed_enrichment_crystallization_necessity priority 3 + scaffolded_sd054_onboarding amend_pending priority 1 (upstream of V3-EXQ-622 anneal-rate amend); Phase 3 telemetry sync_daemon sole-writer hardening + cloud-scaler hub deploy + threshold 3->2 + cloud-1 hub runner re-enabled under Phase 3 co-tenancy + post-result checkout alignment hardening all landed)

• Substrate / governance landings since the 2026-06-01T01:10Z snapshot:
  • Two /governance cycles same day – 16:58Z cycle walked 8 pending experiments with user-approved dispositions: V3-EXQ-614c MECH-341 weakens->non_contributory (instrumentation_defect; confirmed autopsy – score-layer-argmin metric temperature-invariant, C1 mis-specified, C3 substrate PASS 3/3 all arms); V3-EXQ-514k SD-049/SD-015 weakens->non_contributory (ecological wanting/liking dissociation CONFOUNDED by GAP-2 SP-CEM + missing object-bound substrate per failure_autopsy_V3-EXQ-626 routing); V3-EXQ-610b INV-074/MECH-333/MECH-334 non_contributory + substrate_ceiling + pending_retest (D2 control did NOT collapse d2_delta -0.0057; CausalGridWorldV2 test-bed ceiling, identical signature to 610a; dropped mis-tagged MECH-341, marked 610a superseded_by 610b); V3-EXQ-626 superseded by V3-EXQ-626a (Class-1 harness bug – z_goal never driven); V3-EXQ-623 PASS supports MECH-104 (clean discriminative de-commitment: ON n_decommit 24/31 vs ABL 0; supersedes 126; 07:07 empty run superseded by 15:20 canonical). The 17:56Z cycle walked the two remaining items: V3-EXQ-603d FAIL Q-045/MECH-313/MECH-260 non_contributory + substrate_ceiling + pending_retest_after_substrate (LOAD-BEARING autopsy finding: ScaffoldedSD054OnboardingScheduler _train_episode never calls agent.update_z_goal -> z_goal zero-init everywhere -> C4 SUBSTRATE_FAILURE is a Class-1 HARNESS/WIRING ARTIFACT IN A SUBSTRATE MODULE, not a substrate-ceiling falsification); V3-EXQ-625 and V3-EXQ-620 -> evidence_direction=superseded by 625b/620b per /diagnose-errors 17:53Z (gap4 build_config never enabled use_affective_harm_stream, z_harm_a identically 0.0 across all ticks/seeds; vacuous PASS). claims.yaml dated disposition notes appended to MECH-341/INV-074/MECH-333/MECH-334/SD-049/SD-015/Q-045/MECH-313/MECH-260 (validate_claims OK 72 invariants; claims.json 659 entries); manifest edits applied to both flat + run-pack copies (indexer reads run-pack only). substrate_queue: created test_bed_enrichment_crystallization_necessity (priority 3; unblocks INV-074/MECH-333/MECH-334; 610a+610b failure records); amended scaffolded_sd054_onboarding to status=amend_pending (priority 1; strictly UPSTREAM of the V3-EXQ-622 anneal-rate amend; autopsy_ref + amend_hint = wire update_z_goal into scheduler + Stage-0 positive-control gate). pending_review went 1->0 both cycles; closure_drift=0.
  • /diagnose-errors V3-EXQ-625 (headline PASS vs acceptance_pass=false, z_harm_a=0) – ROOT CAUSE code-verified: 625’s headline PASS was vacuous (script defined PASS = ran-cleanly + n>0 distributions, decoupled from the substrate acceptance gate). The real signal acceptance_pass=false had z_harm_a_norm IDENTICALLY 0.0 across all 1027 ticks / 3 seeds – including the 17 external-hazard-event ticks in seed 42. NOT “harm magnitude below the affective noise floor”. The measurement reads agent._current_latent.z_harm_a, which is None unless cfg.latent.use_affective_harm_stream=True. ARM_PHASE1B_BASELINE sets gap4_operating=True -> build_config routed to REEConfig.goal_stream() -> from_dims WITHOUT the SD-011 flags (default False), so the AffectiveHarmEncoder was never instantiated and z_harm_a stayed None (all six consumer-input quantities read exactly 0.0). V3-EXQ-620 axis-a (the result that justified the entire SD-037 axis-b sustained-threat plan) used the identical gap4 path – its pooled-zero distributions were the SAME artifact. Fix: added guarded opt-in build_config(env, arm, enable_affective_harm_stream=True) to goal_pipeline_tier1.py – default False keeps all 12 existing gap4 callers bit-identical; passes limb_damage_enabled through so from_dims auto-sizes harm_obs_a_dim (7 limb-damage / 50 legacy; fixed a 60-vs-17 encoder shape crash). Queued V3-EXQ-620b (axis-a BASELINE env, NO curriculum, stream ON – answers whether axis-b was even necessary) supersedes 620; V3-EXQ-625b (axis-b SD-029 curriculum overlay + stream ON) supersedes 625. Both tie manifest outcome to the substrate-readiness acceptance gate (no more vacuous PASS). claim_ids=[] both (pure-measurement diagnostics; re-evaluated not inherited). Dry-run smoke PASS both: z_harm_a now non-zero (pooled p70 ~0.33 vs old identical 0.0; BLA/CeA receive signal; dacc_pe still 0 = separate deterministic-prediction case). V3-EXQ-620b PASS 2026-06-01T19:01Z confirms axis-a stream now writes – governance flag: once 620b is governance-marked superseder-of-620, reassess whether SD-037 axis-b env-curriculum work was necessary.
  • Five failure-autopsies landed today – (i) V3-EXQ-614c (FAIL, weakens MECH-341, fresh unreviewed FAIL completed 12:45Z after 07:18Z pending_review snapshot): both failing criteria are test-design defects -> recommend MECH-341 non_contributory + supersede + /queue-experiment V3-EXQ-614d. C2 VACUOUS (ARM_1/2/3 bit-identical per seed because reported selected_class_entropy_nats measured at experiment’s own score-layer argmin(last_scores) upstream of the within-class temperature lever; within-class firing diagnostics never recorded). C1 MIS-SPECIFIED (per-seed band [0.72, 0.88] vs 614b cross-seed-MEAN 0.800; ARM_0 per-seed [1.177, 0.530, 0.693] mean=0.7999 = no regression but 0/3 in band). C3 substrate-readiness PASS 3/3 all arms. epistemic_category=instrumentation_defect. (ii) V3-EXQ-610a (FAIL, recovered 05-30, load-bearing INV-074/MECH-333/MECH-334): corrects the rescue session’s conservative weakens read. Governing failure is D2 (negative control did NOT collapse: ARM_0 end_phase_2 1.107 ~= end_phase_3 1.113, d2_delta -0.006); per the script’s OWN pre-registered grid row (c), D2 FAIL = non_contributory. Diagnosis: environment/test-bed ceiling (CausalGridWorldV2 supplies no post-Phase-3 overwriting pressure; 2500 ep makes too-benign-env reading stronger than too-short) -> recommend INV-074/MECH-333/MECH-334/MECH-341 non_contributory + substrate_ceiling + pending_retest_after test-bed enrichment. (iii) V3-EXQ-626 (FAIL, harness bug, not a substrate formation regression): the bespoke _run_episode loop never calls agent.update_z_goal (the only hook into GoalState.update), so z_goal stayed zero-init across all 4 arms; C2/C3 vacuous at 0 < ceiling. Positive controls: 622 S0 PASS (z_goal_peak 0.281/0.439/0.342 same seeds) + V3-EXQ-582a (eff benefit 0.115 > thr). Suspected real abstraction gap: benefit -> write raw z_world into z_goal conflates liking/wanting and location/object; missing object-bound incentive-salience layer (L2-L3). V3-EXQ-514k (dissoc=0.0) is the Class-2 evidence (confounded by GAP-2 SP-CEM). (iv) V3-EXQ-603d (FAIL, scaffolded_sd054_onboarding 4-arm 5th-iter behavioural retest): LOAD-BEARING FINDING – ScaffoldedSD054OnboardingScheduler._train_episode AND 603d’s bespoke P2 never call agent.update_z_goal (zero matches in scaffolded_sd054_onboarding.py) – so z_goal stays at zero-init for every step of every arm; GoalState.update never reached. The manifest’s SUBSTRATE_FAILURE / C4 z_goal=0 label is therefore a Class-1 harness/wiring artifact (same class as the V3-EXQ-626 autopsy) but LIVING IN THE SUBSTRATE MODULE not just a script – NOT substrate-ceiling evidence. This is the true content of 621a’s “z_goal feeding not yet wired correctly” + the 591-triage Section 8 anticipated “prereq (4)”. Second observable: P1 survival gate failed 2/3 seeds (42, 44; only seed 43 to P2; p2_cell_count 5/15, effective N=1), reproducing the 603-chain seed-fragility – but CONFOUNDED by reduced budget (P0/P1=30/30 vs 603c 100/50) + the inert goal pipeline. Biology clear (MECH-313 LC-NE, MECH-260 dACC, vmPFC goal maintenance); no divergence; no claim falsified; all three stay pending_retest_after_substrate / non_contributory; EXQ-445h remains sole valid MECH-260 support. Routing: /implement-substrate AMEND on scaffolded_sd054_onboarding – wire update_z_goal(benefit_exposure, drive_level) into scheduler _train_episode (mirror shared goal_stream_stages_sd054 runner) + Stage-0 positive-control assertion so a z_goal=0 scheduler is unshippable; this is UPSTREAM of the V3-EXQ-622 anneal-rate amend (which assumed update_z_goal was being called); then re-issue 603d -> V3-EXQ-603e via /queue-experiment at restored budget. (v) V3-EXQ-592e (MECH-090 commitment-state release-authority autopsy): score-margin blocks fired (ARM_1=1793, ARM_3=2088) and nav-readiness blocks fired (ARM_2/3=150), but all arms retained hold_rate=1.0 and total_commits=0; per-cell beta totals refine the read. With beta_gate_bistable=True, the score-margin gate is consulted only on not-yet-elevated transition ticks; nav readiness can count blocks on rv-low selections but is still only AND-composed into the elevation call; no readiness-failure path releases an already-elevated beta latch or clears E3 committed state. V3-EXQ-592f proposed as a minimal controlled state-machine probe via real REEAgent.select_action + stubbed E3 SelectionResult, not a full ecological run; V3-EXQ-592f FAIL 2026-06-01T19:43Z confirmed the diagnostic verdict FAIL_NO_RELEASE_AUTHORITY – nav blocks move, beta/e3 occupancy remains 1.0.
  • Phase 3 telemetry sync_daemon sole-writer hardening (ree-v3 main) – documented hub architecture (sync_daemon.phase3_heartbeat_writer owns runner_heartbeats/ + runner_status/ on GitHub; hub runner POST-only via PHASE3_DISABLE_RUNNER_HEARTBEAT_WRITE=1); added deploy/shadow.conf.hub.example; sync_daemon now auto-reverts exclusive telemetry dirt before refusing ticks (fixes the result-writer wedge when only heartbeat paths are dirty); porcelain path parse fix; new contract suite test_phase3_telemetry_dirt_recovery.py; CLAUDE.md + FLEET_CHECKLIST updated.
  • Cloud-scaler hub deploy + cloud-4 surge threshold 3 -> 2 (ree-v3 main 75b23f9) – installed hcloud 1.43.0 amd64 on hub (/usr/local/bin/hcloud), appended HCLOUD_TOKEN (from Mac ~/.config/hcloud/cli.toml context “ree”) to /etc/ree-coordinator.env, copied both unit files to /etc/systemd/system, daemon-reload, systemctl enable --now cloud-scaler.timer (OnCalendar=*:0/5). Verified: dry-run + real systemctl start cloud-scaler.service Result=success ExecMainStatus=0, hub HUB_NAME-skipped, held_by_self veto kept cloud-4 up, surge_threshold=2 loaded. Hub timer is now authoritative 5-min cadence; GHA remains 6-hourly backstop (logic 1:1, both now threshold=2). Note: hcloud install + token + unit install are HUB-LOCAL infra state (not in git).
  • Cloud-1 hub runner re-enabled under Phase 3 co-tenancy – SSH hub: appended PHASE3_DISABLE_RUNNER_HEARTBEAT_WRITE=1 to /etc/systemd/system/ree-runner.service.d/shadow.conf (backup shadow.conf.bak.*); systemctl enable --now ree-runner. Journal confirms hub FILE WRITES skipped gate; coordinator heartbeat fresh (ree-cloud-1 running V3-EXQ-623); sync_daemon still pushing; no local heartbeat churn (stale May-28 file untouched). Docs: CLAUDE.md + FLEET_CHECKLIST D-hub + ree-v3/CLAUDE.md + experiment_runner comment. Hub-only; cloud-2/3/4 unchanged.
  • Post-result checkout alignment hardening (ree-v3 experiment_runner + REE_assembly serve.py, local only) – prepull stash of untracked flat v3 manifests + runner signals before REE_assembly pull; _report_result_and_align runs POST /result then immediate + delayed (45s / 120s) _sync_pull_tick; serve.py replaces ff-only auto-pull with align_ree_assembly_checkout at startup + every 5min (respects TASK_CLAIMS evidence/ skip). Contracts 7/7 PASS.
  • Goal / wanting / liking stream repair intake (REE_assembly master 19d80d5adf + ree-v3 main 16b0c8c) – 7 NEW planning files (docs/thoughts/2026-06-01_goal_wanting_liking_stream_repair.md + thought_intake / literature_synthesis_object_bound_incentive_salience / claim_gap / failure_autopsy_V3-EXQ-626(.md+.json) / goal_stream_repair_diagnostic_ladder_2026-06-01.md) + tests/contracts/test_goalstate_forced_seed_positive_control.py NEW (6/6 PASS). KEY FINDING: V3-EXQ-626 FAIL is a Class-1 HARNESS BUG, not a substrate formation regression – the bespoke _run_episode loop never calls agent.update_z_goal (the only hook into GoalState.update), so z_goal stayed zero-init across all 4 arms; C2/C3 vacuous. Suspected real abstraction gap: benefit -> write raw z_world into z_goal conflates liking/wanting and location/object; missing object-bound incentive-salience layer (L2-L3). 514k (dissoc=0.0) is the Class-2 evidence (confounded by GAP-2 SP-CEM). Flagged: 625 headline PASS vs acceptance_pass=false (harm-consumer-zero); MECH-117 stable vs 514k. Proposals only (PROP-BIND-obj / INCENT-token / CUEWANT / GOALPTR / CONSUME / POSCTRL); NO claims.yaml / queue / ree_core edits.
  • Developmental-window memo (REE_assembly master 12d2a48310) – evidence/planning/goal_pipeline_developmental_window_diagnostic_memo_2026-06-01.md persisted as plan-of-record for V3-EXQ-626/626a; docs/thoughts/2026-06-01_plasticity_window_neuromodulators.md captures the ACh/PV-interneuron/BDNF/state-dependent plasticity-window framing as long-horizon V4-or-late-V3 territory (without committing to substrate-design or new claims); project memory project_plasticity_window_neuromodulators.md registers the territory for future sessions.
• Runner activity since the 2026-06-01T01:10Z snapshot: central evidence/experiments/runner_status.json reports 786 -> 794 cumulative completions (+8); all-time breakdown 195 PASS / 319 FAIL / 87 ERROR / 193 UNKNOWN (deltas PASS +3, FAIL +5, ERROR +0, UNKNOWN +0); last_updated 2026-06-01T01:11:28Z -> 2026-06-01T23:40:21Z – ~1h35m fresh at this read. Today’s eight returns: V3-EXQ-603d FAIL 09:53Z (scaffolded_sd054_onboarding 4-arm 5th-iter; harness/wiring autopsy); V3-EXQ-625 PASS 11:28Z (SD-037 axis-b but vacuous headline – acceptance_pass=false z_harm_a=0; superseded by 625b); V3-EXQ-514k FAIL 11:53Z (SD-049/SD-015 ecological dissociation; confounded; non_contributory); V3-EXQ-614c FAIL 12:45Z (MECH-341 within-class temperature instrumentation defect; non_contributory); V3-EXQ-623 PASS 15:20Z (MECH-104 phasic-spike volatility-interrupt + de-commitment; supports MECH-104, supersedes 126); V3-EXQ-626 FAIL 15:27Z (goal-pipeline 4-arm developmental-window; Class-1 harness bug, z_goal never driven; superseded by 626a); V3-EXQ-620b PASS 19:01Z (SD-037 axis-a stream ON; supersedes 620 vacuous-zero); V3-EXQ-592f FAIL 19:43Z (MECH-090 commitment-state transition authority probe; expected FAIL_NO_RELEASE_AUTHORITY).
• Queue depth: 0 -> 3 items, all claimed at this read – V3-EXQ-610c INV-074 / MECH-333 / MECH-334 post-Phase-3-enrichment crystallization-necessity retest (claimed ree-cloud-3 2026-06-01T18:54Z; supersedes 610b; tests prediction that ARM_0 control NOW collapses under the IGW-023 substrate amend); V3-EXQ-626a goal-pipeline harness-fix re-run (claimed ree-cloud-1 2026-06-01T16:57Z; supersedes 626; P0 positive control is the harness-bug-vs-object-binding adjudicating bit); V3-EXQ-624 ARC-068 / MECH-320 Niv-vs-Salamone opportunity-cost-vs-effort-cost dissociation (claimed DLAPTOP-4.local 2026-06-01T19:58Z; affinity flipped DLAPTOP-4 -> any earlier in day for cloud parallelism, then DLAPTOP-4 claimed after laptop yielded to cloud on 610c / 626a).
• Pending review: 1 -> 0 items maintained across both /governance cycles (pending_review.md regenerated 2026-06-01T18:09Z). V3-EXQ-603d + V3-EXQ-625 intentionally left pending after the 16:58Z cycle for /failure-autopsy and /diagnose-errors respectively; both marked reviewed after the 17:56Z cycle applied dispositions.
• Bottleneck (updated framing): the dominant blocker is now the scaffolded_sd054_onboarding harness / wiring gap identified by V3-EXQ-603d + V3-EXQ-626 failure-autopsies (same epistemic class, different scope). The substrate scheduler _train_episode never calls agent.update_z_goal(benefit, drive) so z_goal stays at zero-init across every step of every arm; the C4 SUBSTRATE_FAILURE classification in 603d’s manifest is a Class-1 harness/wiring artifact LIVING IN THE SUBSTRATE MODULE, NOT a substrate-ceiling falsification. The /implement-substrate AMEND on scaffolded_sd054_onboarding (wire update_z_goal into scheduler + Stage-0 positive-control gate) is strictly UPSTREAM of the V3-EXQ-622 anneal-rate amend and the V3-EXQ-603e re-issue at restored budget. Until that amend lands, the substrate-uniform z_goal-zero pattern across 603 lineage / 626 / 540a-e / 590a / 591 / 598 / 598b remains active. V3-EXQ-626a P0 positive-control gate is the adjudicating bit between harness-bug and object-binding abstraction-gap hypotheses (PASS confirms harness-bug -> object-binding ladder Stage 1+; FAIL contradicts 622 S0 -> /failure-autopsy on GoalState.update). V3-EXQ-610c on ree-cloud-3 (INV-074 / MECH-333 / MECH-334 post-Phase-3-enrichment crystallization-necessity retest) and V3-EXQ-624 on DLAPTOP-4 (ARC-068 / MECH-320 Niv-vs-Salamone) are the next two scientific reads. Governance flag carried forward: once V3-EXQ-620b axis-a PASS is governance-marked superseder-of-620, reassess whether the SD-037 axis-b env-curriculum work was necessary (the original 620 zero artifact underpinned that plan-of-record). The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding.

Immediate Work Queue (This Cycle, 2026-06-02)

1. V3-EXQ-626a outcome adjudication – ree-cloud-1 claimed 16:57Z; P0 positive-control gate is the single bit that adjudicates harness-bug vs object-binding-abstraction-gap. PASS clears the goal-pipeline developmental-window question and routes to object-binding ladder Stage 1+; FAIL contradicts V3-EXQ-622 S0 PASS and routes to /failure-autopsy on GoalState.update.
2. V3-EXQ-610c outcome watch – ree-cloud-3 claimed 18:54Z; tests the prediction that the IGW-20260601-023 InfantCurriculumScheduler Phase 3 enrichment (SD-047 multi_source_dynamics + SD-048 interoceptive_noise + accelerated env_drift) now forces ARM_0 control to collapse (D2 delta >= +0.10) while ARM_1 MECH-333/334 crystallization preserves Phase-2 diversity (D1 delta >= +0.10). If D2 fires, INV-074 / MECH-333 / MECH-334 lift out of pending_retest_after_substrate / non_contributory; if D2 stays flat, the test-bed ceiling is deeper than the Phase 3 enrichment can reach.
3. V3-EXQ-624 outcome watch – DLAPTOP-4 claimed 19:58Z; ARC-068 (action.opportunity_cost_no_op_penalty) Niv-vs-Salamone discriminative test on MECH-320 w_passive implementation. 4-arm Niv x Salamone env x vigor lift design with forced-vigor probe v_t_floor=0.05. C2 dissociation < 0.50 is the load-bearing pass; FAIL routes governance to consider ARC-068 effort-cost-collapse-into-MECH-258 supersession.
4. /implement-substrate AMEND on scaffolded_sd054_onboarding (substrate_queue amend_pending priority 1) – wire agent.update_z_goal(benefit_exposure, drive_level) into ScaffoldedSD054OnboardingScheduler._train_episode (mirror the shared goal_stream_stages_sd054 runner); add Stage-0 positive-control assertion so a z_goal=0 scheduler is unshippable. Strictly UPSTREAM of the V3-EXQ-622 anneal-rate amend.
5. V3-EXQ-603e queueing after substrate amend – re-issue 603d at restored budget (P0/P1 = 100/50 vs 603d’s reduced 30/30) on the hook-fixed scheduler; clears the seed-fragility confound and re-tests Q-045 / MECH-313 / MECH-260 against a substrate that actually drives z_goal.
6. /governance reassessment of SD-037 axis-b necessity (carry-forward) – after V3-EXQ-620b PASS is governance-marked superseder-of-620, reassess whether the SD-037 axis-b env-curriculum work was necessary. The original 620 zero artifact (now confirmed as a stream-OFF measurement artifact, NOT a substrate signal) underpinned the entire axis-b sustained-threat plan-of-record.
7. V3-EXQ-614d (or successor) /queue-experiment (carry-forward from 614c autopsy) – corrected harness measuring committed_classes_p1_counts + within-class diagnostics + cross-seed-mean C1 band; routes the MECH-341 within-class temperature question correctly.
8. V3-EXQ-622 outcome watch (carry-forward) – still in-flight on ree-cloud-1 according to coordinator DB; staged goal-stream S0-S3 four-stage curriculum decomposing the V3-EXQ-621 z_goal failure under the new scaffolded onboarding substrate.
9. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21.

---

Status Snapshot (2026-06-01T01:10Z – nightly docs sync; scaffolded_sd054_onboarding substrate landed (closes behavioral_diversity_isolation:GAP-C prereq (2); new experiment-harness scheduler at experiments/scaffolded_sd054_onboarding.py with three-phase P0/P1/P2 curriculum (frozen goal pipeline + reef-half spawn -> linear-anneal hazard/proximity_harm/drive-to-fire/z_beta-threshold -> target env frozen-policy measurement); new env kwarg reef_bipartite_agent_spawn_in_reef_half on CausalGridWorldV2; master switch use_scaffolded_sd054_onboarding_scheduler default False; 14 phase-config knobs; 645/645 contracts + 17 new contracts PASS; bit-identical OFF guarantee verified; V3-EXQ-621 ERROR superseded by V3-EXQ-621a PASS at 2026-05-31T23:09Z with emit_outcome + per-cell p1_episode_lengths); SD-056 multi-step rollout stability amend landed (ree-v3 d327b89; multi-step contrastive h=5 + per-step output norm clamp ratio=2.0; V3-EXQ-617 substrate-readiness PASS); InfantCurriculumScheduler Phase 0->1 H_pos floor recalibration landed (closes behavioral_diversity_isolation:GAP-C prereq (3)); MECH-341 cluster autopsy (614b + 615 + 616 convergent reading – ARC-065 supports promotion-eligibility candidate->provisional, MECH-341 non_contributory score-layer preserver, Q-054 mixed definitive negative answer to scale-lever); flat-vs-runs propagation-failure mirror campaign cleared 11 claims from active conflicts table 95->84 (ARC-045/INV-010/MECH-166/MECH-261/MECH-302/MECH-320/SD-017/SD-033a/Q-045/MECH-313/MECH-260 + ARC-065 ratio 0.231->0.043); cloud-scaler GHA->hub systemd migration prepared (ree-v3 d641419, deploy operator-driven); /governance cycle opened 2026-05-31T19:08Z still active at this read; central runner_status freshness live at this read (786 cumulative, +13 since 2026-05-31T01:10Z; 192 PASS / 314 FAIL / 87 ERROR / 193 UNKNOWN); pending_review.md 1->0 (all walked via active /governance cycle); queue items[] now empty (V3-EXQ-621a + V3-EXQ-622 runner-claimed))

• Substrate / governance landings since the 2026-05-31T01:10Z snapshot:
  • scaffolded_sd054_onboarding substrate (curriculum.scaffolded_sd054_onboarding, 2026-05-31, ree-v3 main 28ebd3d) – closes behavioral_diversity_isolation:GAP-C prereq (2) substrate landing (Cluster B / V3-EXQ-603c substrate-uniform z_goal-zero family addressed by the V3-EXQ-490g-cohort autopsy 2026-05-29). Three-phase scheduler at experiments/scaffolded_sd054_onboarding.py (NEW; experiment-harness layer alongside infant_curriculum.py + committed_mode_curriculum.py precedent; ree_core/ otherwise UNTOUCHED). P0: frozen goal pipeline (use_mech295_liking_bridge=False + use_mech307_conjunction=False runtime mutation) + reef-half spawn admissibility (new env kwarg reef_bipartite_agent_spawn_in_reef_half default False on CausalGridWorldV2; pool predicate widened to agent_band OR reef_half when True) + sub-target proximity_harm_scale 0.05 + relaxed hazard density; E1+E2 training over 30 episodes at 200 steps/episode. P1: linear anneal hazard_food_attraction 0.0 -> 0.7, proximity_harm_scale 0.05 -> 0.1, mech295_min_drive_to_fire 1.0 -> 0.01, mech307_conjunction_z_beta_threshold 0.6 -> 0.3; spawn admissibility narrows back to midline band; goal pipeline UNFROZEN; end-of-P1 survival gate (median episode length >= 75 over last 10 episodes, Fix D retained from V3-EXQ-603c). P2: target env config (hazard_food_attraction=0.7, proximity_harm_scale=0.1, num_hazards=4, num_resources=5; matches V3-EXQ-603b GAP-4 Tier-1 measurement env); policy frozen (no optimizer steps); measures z_goal_norm_peak, approach_commit_rate, bridge_cue_fires, dacc_bias_nonzero_steps per episode. Master switch use_scaffolded_sd054_onboarding_scheduler default False on ScaffoldedSD054OnboardingConfig (NOT surfaced through REEConfig.from_dims); 14 phase-config knobs match memo Config Surface table. 645/645 contracts + 17 new scaffolded_sd054_onboarding contracts PASS; bit-identical OFF guarantee verified. MECH-094: N/A (waking-stream env + agent state; no simulation / replay write surface). V3-EXQ-621 ERROR (runner sentinel misclass + missing emit_outcome; manifest recovered from ree-cloud-3) superseded by V3-EXQ-621a PASS 2026-05-31T23:09Z with emit_outcome + P1 survival diagnostics (per-cell p1_episode_lengths[] + verdict lines); V3-EXQ-622 staged goal-stream S0-S3 decomposing the 621 z_goal failure into a four-stage curriculum queued same evening.
  • SD-056 multi-step rollout stability amend (e2.world_forward.multi_step_contrastive, 2026-05-31, ree-v3 d327b89) – amend to the 2026-05-29 SD-056 landing: multi-step contrastive horizon h=5 (vs h=1) + per-step output norm clamp ratio=2.0 prevents multi-step rollout drift; resolves the V3-EXQ-569a self-anchored-targets NaN cluster + provides substrate-readiness preconditions for the 614 lineage behavioural validation chain. V3-EXQ-617 substrate-readiness PASS 2026-05-31T11:31Z confirming amend stability + V3-EXQ-614a/b multi-arm rollout integrity (zero NaN/Inf across 162k steps at ARM_2 ALL_ON; 614 lineage entropy 0.684 -> 0.800 nats post-amend).
  • InfantCurriculumScheduler Phase 0->1 H_pos floor recalibration (2026-05-31) – closes behavioral_diversity_isolation:GAP-C prereq (3) Phase 0->1 exit signal recalibration. Sibling work to scaffolded_sd054_onboarding; together they close prereqs (2) and (3) for the goal-pipeline default-config z_goal generation gate. 645/645 contracts + 7/7 preflight PASS.
  • MECH-341 cluster autopsy (V3-EXQ-614b + V3-EXQ-615 + V3-EXQ-616 convergent reading; sister artifact to the 2026-05-29 608+611 cluster autopsy; the 614b single-target autopsy stays intact). V3-EXQ-615 09:31Z PASS supports ARC-065 – ARM_2 ALL_ON 4.33 unique classes / 1.111 nats vs ARM_0 BASE_OFF + ARM_1 MATCHED_NOISE both collapsing to single class 0.0 nats; MECH-313 noise_floor alone cannot reproduce distributed-pathway diversity. V3-EXQ-616 FAIL_no_floor_under_max_swept_scale on MECH-341 B_only isolation across {1.0, 2.0, 4.0, 8.0} entropy_bias_scale – BIT-IDENTICAL per-seed results at every scale (frac_pre_ge2=0.0 / n_unique=1 / entropy=0.0 universally; mathematical proof that uniform additive entropy bias of any magnitude in [1.0, 8.0] cannot move a single-class CEM proposer output). V3-EXQ-614b substrate-coupling FAIL under SD-056-amended substrate (ARM_2 ALL_ON highest absolute entropy of any 614 run at 0.800 nats but MECH-341 marginal contribution shrank to 0.087 vs 0.100 C2 threshold; NOT a falsification – upstream cluster now does more diversity work). Per-claim direction overrides recommended for /governance: ARC-065 supports (clear pending_retest_after_substrate; promotion-eligibility candidate -> provisional surfaced); MECH-341 non_contributory (claim correctly characterised as score-layer preserver, not in-isolation diversity generator); Q-054 mixed (definitive negative answer to scale-lever framing). Process incidents surfaced: V3-EXQ-615 same-queue_id dual-manifest (08:22Z dry-run smoke + 09:31Z canonical PASS; manifest supersession applied 2026-05-31T19:04Z); C2 necessity_delta_threshold=0.1 in 614 lineage scripts is stale under SD-056-amended substrate.
  • Multi-round flat-vs-runs propagation-failure mirroring – three rounds; 29 mirrors total. Root cause: build_experiment_indexes.py:294 globs **/runs/**/manifest.json only; reclassifications applied to top-level flat JSON were indexer-invisible. Round 1 (V3-EXQ-573 canonical fix) cleared MECH-314 entirely. Round 2 mirrored 28 audited c.1+c.2 candidates split into SAFE (17: flat commit newer than runs + explicit governance commit subject) + SAME (11: single governance commit touched both files but only flat reclassified) + DANGER (1: V3-EXQ-603 auto-sync overwrote runs after morning reclassification, deferred). Round 3 mirrored the V3-EXQ-603 DANGER entry on fresh-governance review (autopsy facts airtight, sibling runs already non_contributory on both flat+runs, runs manifest CREATED by auto-sync not overwritten). Active conflicts table: 95 -> 84 claims (11 dropped); fully cleared ARC-045 / INV-010 / MECH-166 / MECH-261 / MECH-302 / MECH-320 / SD-017 / SD-033a / Q-045 / MECH-313 / MECH-260; ARC-065 weakens 7 -> 1, ratio 0.231 -> 0.043 (huge cleanup downstream of round 1 + 5 weakens cleared from EXQ-569 + EXQ-572 governance reclassification). Root-cause recommendation re-surfaced: the reclassify workflow should write the canonical runs//manifest.json as primary; flat JSON should auto-sync write-through or be dropped from the schema.
  • cloud-scaler GHA -> hub systemd migration prepared (ree-v3 d641419) – coordinator/deploy/cloud-scaler.{py,service,timer,md} adds a hub-resident OnCalendar=*:0/5 oneshot replacing the every-15-min GHA workflow whose actual gaps had grown to 60-273 minutes under GHA load. Python is a 1:1 port of the bash decision matrix; HUB_NAME guard (cloud-1 skip), HELD_BY_SELF veto (2026-05-30 fleet-incident invariant), surge sister-state pre-check, HEARTBEAT_FRESH_MIN >= 35 floor all preserved. Adversarial parity review 10/10 invariants OK. GHA workflow downgraded to cron 0 */6 * * * as hub-down backstop. Deploy is operator-driven; not deployed this session.
  • V3-EXQ-615 dual-manifest supersession (ree-v3-cluster process flag (a) resolved 2026-05-31T19:04Z) – the 08:22Z manifest was a dry_run=True smoke-budget artifact emitted to flat evidence/experiments/; the 09:31Z is the canonical full-scale run (3 seeds, 200 steps/ep, ree-cloud-2 PASS). Applied CLAUDE.md supersession pattern: evidence_direction=superseded + evidence_direction_note on BOTH the flat and the canonical runs/<run_id>/manifest.json. ARC-065 scoring delta verified pre/post rebuild: direction_counts {supports:46, mixed:5, weakens:1} -> {supports:46, mixed:5, weakens:0}; experimental_confidence 0.899 -> 0.974.

  • SD-037 axis (a) consumer-input recalibration plan landed (evidence/planning/sd_037_axis_a_consumer_input_recalibration_plan.md, 2026-05-31T16:54Z) – four-phase plan for V3-EXQ-483e FAIL diagnosis (BLAConfig.arousal_threshold_on=0.4 / CeAConfig.fast_route_threshold=0.5 / PAGFreezeGateConfig.theta_freeze=2.0 + duration_input_threshold=0.4 / dACC PE-magnitude floors all sit above fishtank baseline signal magnitudes; broadcast amplifies zero baseline outputs). Phase 1 substrate-readiness diagnostic V3-EXQ-620 PASS 2026-05-31T18:36Z logged the six raw consumer-input distributions (z_harm_a_norm,

    LowFreq(z_harm_a)

    , sustained-product, PE) over fishtank baseline; closes plan Phase 1. Phase 2 deterministic p70 recalibration rule + sensitivity p60/p80 fallback queued.

  • V3-EXQ-618 SD-049 Phase 3 SD-032 consumer cascade substrate-readiness PASS (2026-05-31T17:59Z) – 3-arm Phase A (ARM_0 BASELINE_CASCADE_OFF / ARM_1 CASCADE_ON_COMBINER_ONLY / ARM_2 CASCADE_ON_AXIS_AWARE with per-episode _current_goal_axis_idx cycling) end-to-end smoke + Phase B direct-API probes of AIC.tick (per-axis vs None) and MECH295LikingBridge.compute_approach_cue_score_bias (axis-aware vs combiner-only at non-max axis). 4/4 acceptance criteria PASS. PASS routes to V3-EXQ-619 load-bearing behavioural validation (V3-EXQ-514g-successor structure with trained agent + active goal_state; unblocks goal_pipeline:GAP-2 cluster).
  • IGW-038..042 cohort proposal-truth-up – 5 proposals (EXP-0068 SD-047 / EXP-0076 ARC-018 / EXP-0078 MECH-314 / EXP-0080 MECH-318 / EXP-0081 SD-011) flipped to status=executed (4) or status=skipped (1) after audit confirmed why_now=active_conflict tag was stale (all 5 target claims at conflict_ratio < 0.3, SD-011 already stable). Workset generator active_conflict signal flagged as stale across this cluster.
• Runner activity since the 2026-05-31T01:10Z snapshot: central evidence/experiments/runner_status.json reports 773 -> 786 cumulative completions (+13); all-time breakdown 192 PASS / 314 FAIL / 87 ERROR / 193 UNKNOWN (deltas PASS +7, FAIL +3, ERROR +3, UNKNOWN +0); last_updated 2026-05-30T18:47:00Z -> 2026-06-01T01:11:28Z – live at this read. Today’s ten returns: V3-EXQ-569d PASS 05:36Z (ARC-065 / MECH-341 floor-recalibrated falsifier); V3-EXQ-519b PASS 06:59Z (SD-051 / MECH-304 conditioned-safety-store readiness retest after MECH-302 gate (c) lifted); V3-EXQ-615 PASS 09:31Z (ARC-065 Rung-1 matched-entropy control); V3-EXQ-617 PASS 11:31Z (SD-056 multi-step rollout stability amend substrate-readiness); V3-EXQ-616 FAIL 14:15Z (Q-054 entropy_bias_scale sweep – mathematical proof MECH-341 isolation structurally not reachable via the score-layer scale lever); V3-EXQ-618 PASS 17:59Z (SD-049 Phase 3 SD-032 consumer cascade substrate-readiness); V3-EXQ-614b FAIL 18:20Z (MECH-341 P3 behavioural falsifier under SD-056-amended substrate – substrate-coupling FAIL, NOT a falsification per autopsy); V3-EXQ-620 PASS 18:36Z (SD-037 axis (a) Phase 1 consumer-input distributions diagnostic); V3-EXQ-621 ERROR 20:23Z (scaffolded_sd054_onboarding substrate-readiness; manifest recovered from cloud-3); V3-EXQ-621a PASS 23:09Z (scaffolded_sd054_onboarding substrate-readiness with emit_outcome + P1 survival diagnostics).
• Queue depth: 3 -> 0 items at items[] read – the V3-EXQ-490j + V3-EXQ-519b + V3-EXQ-569d entries drained during the day (519b + 569d returned PASS; 490j runner-claimed). V3-EXQ-622 staged goal-stream S0-S3 (queued 2026-05-31T21:23Z with priority 35 four-stage curriculum decomposing the V3-EXQ-621 z_goal failure) was runner-claimed promptly so it does not appear at this read. V3-EXQ-620 SD-037 Phase 1 distributions diagnostic (queued 17:43Z, PASS 18:36Z) drained same day.
• Pending review: 1 -> 0 items (pending_review.md regenerated 2026-05-31T19:25Z; last review 2026-05-31T19:15Z). All walked via the active /governance cycle opened 2026-05-31T19:08Z. V3-EXQ-490i FAIL applied. The V3-EXQ-617 / V3-EXQ-619 / V3-EXQ-621a substrate-readiness manifests are discussed via discussed_experiment_dirs (scoring_excluded).
• Bottleneck (updated framing): the goal-pipeline / training-regime substrate enrichment is now the load-bearing constraint behind the substrate-uniform monomodal-V_s monostrategy tail across 483c / 524a / 603 lineage / 540a-e / 590a / 591 / 598 / 598b. The Layer-A E2 forward-model collapse fix (SD-056 + multi-step amend) is substrate-readiness validated (V3-EXQ-617 PASS); the scaffolded_sd054_onboarding substrate is now landed (V3-EXQ-621a PASS provides the first substrate-readiness signal). The next-cycle bottleneck is the V3-EXQ-622 staged goal-stream S0-S3 outcome watch – decomposes the V3-EXQ-621 z_goal failure into a four-stage curriculum (S0 frozen-goal warmup -> S1 sparse goal pipeline -> S2 dense goal pipeline -> S3 target env frozen-policy measurement) to localise where the substrate-uniform monostrategy collapse re-emerges under the new scaffolded onboarding. The MECH-341 cluster autopsy + the flat-vs-runs mirror campaign together cleared 11 claims from the active conflicts table (95 -> 84) and surfaced ARC-065 promotion-eligibility candidate -> provisional for the next governance walk. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding – not closed today.

Immediate Work Queue (This Cycle, 2026-06-01)

1. V3-EXQ-622 outcome watch – staged goal-stream S0-S3 curriculum (priority 35; four-stage decomposition of the V3-EXQ-621 z_goal failure under the new scaffolded onboarding substrate; localises where the substrate-uniform monostrategy collapse re-emerges). PASS at each stage routes to deeper behavioural validation; FAIL routes to per-stage failure-autopsy.
2. V3-EXQ-621a outcome interpretation – scaffolded_sd054_onboarding substrate-readiness PASS provides first signal but z_goal_norm_peak / approach_commit_rate / bridge_cue_fires / dacc_bias_nonzero_steps need to be cross-checked against the P2 frozen-policy measurement gate; if the substrate clears the prereq (2) acceptance criterion (z_goal_norm_peak >= 0.1 on >= 2/3 seeds in at least one arm OR cascade behaviourally consequential), the behavioural cluster validation V3-EXQ-603d / 591b is queueable.
3. /governance cycle continuation (opened 2026-05-31T19:08Z, still active) – apply per-claim direction overrides from MECH-341 cluster autopsy (ARC-065 supports + promotion eligibility, MECH-341 non_contributory, Q-054 mixed); apply ARC-065 scoring delta from V3-EXQ-615 dual-manifest supersession (experimental_confidence 0.899 -> 0.974); review the 11 claims dropped from active conflicts for promotion eligibility.
4. V3-EXQ-619 SD-049 Phase 3 behavioural validation queueing decision – gated on V3-EXQ-618 PASS (now landed). The behavioural validation has been verbally reserved for V3-EXQ-619 with trained agent + active goal_state; queue after V3-EXQ-618 sentinel routing confirms readiness.
5. SD-037 axis (a) Phase 2 recalibration – after V3-EXQ-620 distributions diagnostic PASS, compute per-knob 70th-percentile recalibration rule for BLAConfig.arousal_threshold_on / CeAConfig.fast_route_threshold / PAGFreezeGateConfig.duration_input_threshold + theta_freeze / DACCConfig.dacc_precision_scale as per-experiment overrides; Phase 3 verification diagnostic to follow; Phase 4 V3-EXQ-483f via /queue-experiment only after Phase 3 acceptance.
6. /diagnose-errors SD-032b dACC bundle -> E3 score_bias adapter (carry-forward) – the secondary finding from V3-EXQ-490i autopsy (orthogonal SD-032b consumer-pathway wiring gap; dacc_bias_nonzero_steps=0 across all ARM_1 seeds despite cfg.use_dacc=True). Same wired-but-inert signature as the SD-037 489d ceiling.
7. cloud-scaler hub systemd deploy decision – coordinator/deploy/cloud-scaler.{py,service,timer,md} prepared but not deployed; operator decision on cutover timing (GHA workflow still active with 6h cron as backstop).
8. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21.

---

Status Snapshot (2026-05-31T01:10Z – nightly docs sync; SD-022 scheduled-injection extension landed today (env-side curriculum that injects damage directly into self.limb_damage independent of agent action or hazard contact; supplies detectable damage->heal trajectories for MECH-302 SufferingDerivativeComparator regardless of trained avoidance policy; 5 env-only kwargs NOT surfaced through REEConfig.from_dims; 565/565 contracts PASS; V3-EXQ-517c PASS 2/3 ARM_A seeds 160.3 events/seed + 3/3 ARM_B seeds 0 events; cleared MECH-302 + MECH-303 v3_pending gates via IGW-021 17:16Z and lifted gate (c) for MECH-304 V3-EXQ-519 conditioned-inhibition); SD-037 consumer-cascade MECH-281 motor-coupling axis amend (4 additional override_signal consumer sites wired – LateralPFCAnalog eff_eta scaling + BLAAnalog encoding_gain scaling + CeAAnalog mode_prior+fast_prime amplification + BetaGate urgency_interrupt threshold attenuation; all 4 scalar gains default 0.0 bit-identical OFF; 556/556 contracts + 13 new MECH-281 contracts PASS; V3-EXQ-483e queued for validation); runner stack hardening (SIGTERM phantom-completion fix _transient_exit_codes extended to {137, -9, -11, -15, 143} + sentinel-detection bare outcome regex RE_BARE_OUTCOME added + Phase 3 coordinator /claim claim_log INSERT landed first row since 2026-05-21 + heartbeat write-gate scope corrected COORDINATOR_TIMEOUT 3->10s); IGW housekeeping batch (4 wrong-route proposals gated + auto-spawn loop root-caused at generate_inter_governance_workset.py _substrate_resolved and patched + workset regen 53/20/3 ready/in_flight); central runner_status freshness ~6.5h at this read (773 cumulative, +5 since 2026-05-30T01:10Z; 185 PASS / 311 FAIL / 84 ERROR / 193 UNKNOWN); pending_review.md 2 -> 1 (V3-EXQ-490i FAIL autopsied + V3-EXQ-483d cleared per autopsy applied this morning; V3-EXQ-612b cleared after diagnose-errors session); queue 1 -> 3 (V3-EXQ-490j MECH-295 successor + V3-EXQ-519b SD-051/MECH-304 readiness + V3-EXQ-569d ARC-065/MECH-341 floor-recalibrated; V3-EXQ-569e mechanism probe queued same day and runner-claimed))

• Substrate / governance landings since the 2026-05-30T01:10Z snapshot:
  • SD-022 scheduled-injection extension (environment.scheduled_limb_damage_curriculum) – env-side curriculum that periodically injects damage directly into self.limb_damage independent of agent action or hazard contact, supplying detectable damage->heal trajectories so the MECH-302 SufferingDerivativeComparator (SD-050) has reliable suffering signals regardless of a trained avoidance policy. Triggered by failure_autopsy_V3-EXQ-517b_2026-05-30: three FAIL discriminative-pair attempts (V3-EXQ-517 / 517a / 517b, 2026-05-04..06) ruled out parameter tuning – trained avoidance policy filters out hazard-contact -> heal trajectories the comparator needs. Architecturally orthogonal to SD-029 scheduled_external_hazard: SD-029 relocates a hazard adjacent to the agent (still requires agent contact); SD-022 scheduled-injection bypasses contact entirely (allostatic / externally-imposed tissue insult). Five new env-only kwargs (NOT surfaced through REEConfig.from_dims; matches SD-022 / SD-023 / SD-029 / SD-047 / SD-048 / SD-049 / SD-054 precedent): scheduled_limb_damage_enabled (False), _interval (50), _prob (0.5), _magnitude (0.4), _limb_selection ("random" or "all"). Preconditions: enabled=True requires limb_damage_enabled=True (loud-not-silent ValueError). Always-present info dict tags. ML/AI anchor: Bengio 2009 automated curriculum learning – stochastic gate + random limb selection mitigate schedule-prediction degenerate solutions. MECH-094 N/A (env observation stream). 565/565 contracts + 7/7 preflight PASS. V3-EXQ-517c PASS 2026-05-30T12:45Z (2/3 ARM_A seeds 160.3 events/seed + 0 ARM_B); cleared MECH-302 + MECH-303 v3_pending gates via IGW-021 17:16Z and lifted gate (c) for MECH-304 V3-EXQ-519 conditioned-inhibition.
  • SD-037 consumer-cascade (MECH-281 motor-coupling axis amend) (regulators.broadcast_override.consumer_cascade) – amend session (NOT a fresh SD landing) triggered by V3-EXQ-483d FAIL (2026-05-29) substrate-ceiling diagnosis: with GoalState seeding + PAG freeze-gate consumers already wired (2026-04-25) but SalienceCoordinator slot dormant in the validation env and PFC/BLA/CeA/beta-gate sites unwired, override_signal had nowhere to land where it would move goal_norm_peak against the MECH-295 bridge baseline. Four additional consumer sites wired (all gated by 0.0-default scalar gains – bit-identical OFF): (i) LateralPFCAnalog (SD-033a) eff_eta scaled by 1 + override_eta_gain * override_signal (orexin-recruited state accelerates rule_state EMA); (ii) BLAAnalog (SD-035) encoding_gain scaled by 1 + override_encoding_gain * override_signal (Roozendaal 2011 orexin -> NE / amygdala enhanced LTP); (iii) CeAAnalog (SD-035) mode_prior + fast_prime scaled by 1 + override_amplitude_gain * override_signal (re-clipped to mode_prior_log_odds_max); (iv) BetaGate / MECH-091 urgency_interrupt_threshold attenuated by max(0, 1 - override_beta_interrupt_gain * override_signal) (orexin escape-from-freeze on motor side, parallel to PAG alpha_override). Four new REEConfig + from_dims knobs: override_pfc_eta_gain, override_bla_encoding_gain, override_cea_amplitude_gain, override_beta_interrupt_gain (all default 0.0). MECH-094 preserved via existing simulation_mode arguments on BLA / CeA tick paths. 556/556 contracts + 13 new MECH-281 contracts PASS. V3-EXQ-483e queued (4-arm successor under 483 lineage; claim_ids=[SD-037, MECH-280, MECH-281]; re-runs 483d ARM config with use_salience_coordinator=True + all four consumer-cascade gains>0 + PAG-engaging env via SD-036+MECH-279 freeze-engaging substrate).
  • MECH-302 + MECH-303 v3_pending cleared (IGW-021 at 17:16Z) on V3-EXQ-517c PASS. substrate_queue MECH-302 status implemented -> validated; gate (c) for MECH-304 V3-EXQ-519 conditioned-inhibition experiment lifted; substrate_queue MECH-302 + MECH-303 entries updated with validation evidence + cleared blocker text.
  • Runner stack hardening – four parallel runner-side improvements: (i) SIGTERM phantom-completion fix (ree-v3 main c8288f1) extends _transient_exit_codes to {137, -9, -11, -15, 143}; cloud-scaler shutdowns now intercept as infra-crash + release_active_claim + leave queue entry intact rather than silently writing phantom completion rows that then trip preflight on next boot; 5 new contract tests in tests/contracts/test_runner_sigterm_no_phantom_completion.py. (ii) Sentinel-detection bare outcome regex (ree-v3 main 9c187d0) – added RE_BARE_OUTCOME = re.compile(r'(?im)^outcome:\s+(PASS|FAIL)\b') alongside the existing six sentinel patterns; scripts emitting bare outcome: PASS/FAIL lines (no Done. prefix, no [EXQ-x] prefix) now classify correctly. Root cause of V3-EXQ-517c initial ERROR mis-classification (105-min PASS run logged ERROR despite valid manifest). (iii) Phase 3 coordinator /claim endpoint claim_log INSERT (ree-v3 main 0128cdc) – coordinator/app.py /claim non-shadow branch now calls db.log_claim(...) after db.try_claim; was silently dropping all claim_log writes since the 2026-05-21 Phase 3 cutover replaced the shadow path (which logged) with the authoritative try_claim path (which didn’t). First claim_log row landed since 2026-05-21. (iv) Heartbeat write-gate scope correction + COORDINATOR_TIMEOUT 3 -> 10s (ree-v3 main d82af98) – runner_remote_control.write_heartbeat PHASE3_DISABLE_RUNNER_HEARTBEAT_WRITE gate now scopes to local file write only; the coordinator POST always fires (was suppressing both before, breaking the canonical materialiser path because phase3_heartbeat_writer’s WHERE clause requires heartbeat_payload_json IS NOT NULL); COORDINATOR_TIMEOUT default raised 3 -> 10s for WireGuard-payload safety. 5 new contracts in tests/contracts/test_runner_heartbeat_write_gate_preserves_coordinator_post.py.
  • IGW housekeeping batch – 4 wrong-route proposals flipped to status=gated (EXP-0003 MECH-334 / EXP-0051 ARC-045 / EXP-0062 Q-045 / EXP-0064 MECH-166); IGW auto-spawn loop root-caused at REE_assembly/scripts/generate_inter_governance_workset.py _substrate_resolved and patched (field-precedence error – implementation_status OR status made first-truthy win, so implementation_status=implemented_env_curriculum_amend shadowed status=validated and the entry never resolved; new _status_resolved() helper consults both fields and accepts the literals + implemented_* prefix + phase_N_implemented regex while excluding *_pending_*); 12-case unit test PASS covering all real-world status strings; workset 53 -> 48 items after generator fix; status carry-forward extended in build_experiment_indexes.py line 4084 to include gated metadata (gated_at_utc + gated_by_session + gating_reason + predecessor_disposition + release_condition + superseded_by); /inter-governance-brief regenerated workset (53 items; 20 ready; 0 pending_user; 3 in_flight).
  • Three /failure-autopsy artifacts – V3-EXQ-490h NO-OP close (already autopsied 06:22Z under the 3-target manifest-pipeline cluster); V3-EXQ-490i MECH-295 GAP-4 Tier-1 (bridge sign-test PASSES; metric-design contamination on C3_lift_vs_baseline + orthogonal C2 dACC wiring gap; recommended MECH-295 narrow_supports); V3-EXQ-569c SD-056 + ARC-065 / MECH-341 (C1 borderline FAIL + strong C3 entropy lift; floor-calibration miss not substrate weakness; recommended ARC-065 supports + MECH-341 supports).
• Runner activity since the 2026-05-30T01:10Z snapshot: central evidence/experiments/runner_status.json reports 768 -> 773 cumulative completions (+5); all-time breakdown 185 PASS / 311 FAIL / 84 ERROR / 193 UNKNOWN (deltas PASS +2, FAIL +3, ERROR +0, UNKNOWN +0); last_updated 2026-05-29T20:17:10Z -> 2026-05-30T18:47:00Z – ~6.5h fresh at this read. Today’s three returns: V3-EXQ-517c PASS 2026-05-30T12:45Z on the SD-022 scheduled-injection curriculum (cleared MECH-302 + MECH-303 v3_pending and lifted gate (c) for MECH-304 V3-EXQ-519); V3-EXQ-569c FAIL 2026-05-30T06:00Z on the SD-056 + ARC-065 / MECH-341 matched-entropy FP-2 falsifier (C1 borderline + strong C3 entropy lift; autopsy upgrade supports); V3-EXQ-490i FAIL 2026-05-30T14:31Z on MECH-295 GAP-4 Tier-1 (bridge sign-test PASSED 3/3 + metric-design contamination on C3_lift + dACC wiring gap; autopsy upgrade narrow_supports).
• Queue depth: 1 -> 3 items – V3-EXQ-592c V3-EXQ-612d drained; V3-EXQ-483d / 612b reviewed via the morning governance walk. New entries: V3-EXQ-490j (MECH-295 successor, severed-bridge baseline with z_goal_enabled=False ARM_0 + direct bridge-magnitude probe replacing contaminated goal_norm_peak delta; priority 350), V3-EXQ-519b (SD-051 / MECH-304 conditioned-safety-store readiness; priority 340), V3-EXQ-569d (ARC-065 / MECH-341 floor-recalibrated falsifier; priority 310). V3-EXQ-569e MECH-immediate Pathway-A-vs-B mechanism probe was queued 17:13Z (priority 305) and runner-claimed earlier this evening so it no longer appears in items[] at this read.
• Pending review: 2 -> 1 item (pending_review.md regenerated 2026-05-30T19:42Z) – V3-EXQ-490i FAIL already autopsied per /failure-autopsy session 19:48Z and awaiting governance application of the recommended per-claim direction shift. V3-EXQ-483d failure-autopsy applied this morning per the 18:20Z /governance walk; V3-EXQ-612b ERROR cleared via the diagnose-errors session also applied at 18:20Z. V3-EXQ-517c PASS + V3-EXQ-569c FAIL both walked via the /governance cycle.
• Bottleneck note (unchanged framing; today’s amends close substrate-ceiling pockets but Layer-A E2 fix still validating): SD-056 lever-B InfoNCE contrastive next-state landed yesterday (2026-05-29) closes the upstream E2 world-forward per-candidate z_world collapse at the substrate level; V3-EXQ-613 substrate-readiness PASS confirmed the diagnostic, and V3-EXQ-569b/c/d/e plus the scaffolded_sd054_onboarding substrate-queue entry are the load-bearing follow-on validation chain. Today’s two amends (SD-022 scheduled-injection + SD-037 consumer-cascade) closed separate substrate-ceiling pockets (MECH-302 + MECH-281) but neither removes the Layer-A (E2 forward-model collapse) cause flagged as the root structural blocker by the V3-EXQ-571 investigation. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding – not closed today.

Immediate Work Queue (This Cycle, 2026-05-31)

1. V3-EXQ-490j outcome watch – MECH-295 cascade Tier-1 successor with TRUE severed-bridge ARM_0 (z_goal_enabled=False) + direct bridge-magnitude probe (per-tick anticipatory liking-stream write + per-candidate approach-cue score_bias magnitude + bridge_cue_fires + bridge_write_fires) replacing the contaminated goal_norm_peak delta + consistent total-step budget across seeds. Tests whether the bridge intactness produces a measurable per-tick magnitude lift independent of the metric-design contamination flagged by the V3-EXQ-490i autopsy.
2. V3-EXQ-569d outcome watch – ARC-065 / MECH-341 floor-recalibrated falsifier successor (supersedes V3-EXQ-569c). Tests whether the substrate-operative ~0.045 pairwise_dist measured on the 569c FAIL clears the recalibrated 0.03 floor while preserving the strong C3 entropy lift; PASS would upgrade ARC-065 + MECH-341 to supports formally.
3. V3-EXQ-569e outcome watch – ARC-065 / MECH-341 Pathway-A-vs-B mechanism probe (parallel chip, not successor). Dissociates whether the strong C3 entropy lift on 569c came from Pathway A (per-candidate z_world variance propagation through E3 softmax routing) vs Pathway B (E2 rollout dynamics shift producing different scoring landscape). Routes the next substrate-revisit decision.
4. V3-EXQ-519b outcome watch – SD-051 / MECH-304 conditioned-safety-store readiness retest with the MECH-302 gate (c) now lifted. Tests whether the conditioned-inhibition pipeline produces the expected substrate-readiness signal on the rescued substrate; PASS would clear MECH-304 v3_pending.
5. V3-EXQ-483e outcome watch – SD-037 consumer-cascade behavioural validation (4-arm successor under 483 lineage). Tests whether the four new consumer sites + use_salience_coordinator=True + PAG-engaging env via SD-036+MECH-279 freeze-engaging substrate restore goal_norm_peak lift across the broadcast_override axis (vs bit-identical action_counts on V3-EXQ-483d).
6. /diagnose-errors SD-032b dACC bundle -> E3 score_bias adapter – the secondary finding from the V3-EXQ-490i autopsy (orthogonal SD-032b consumer-pathway wiring gap; dacc_bias_nonzero_steps=0 across all ARM_1 seeds despite cfg.use_dacc=True). Same wired-but-inert signature as the SD-037 489d ceiling; user-confirmed scope was to route as a separate /diagnose-errors session rather than fold into the MECH-295 successor.
7. /implement-substrate scaffolded_sd054_onboarding session-of-record (carry-forward) – the substrate-implementation session for the priority-1 substrate_queue entry created by yesterday’s evening governance cycle (unblocks 9 claims). Design memo at evidence/planning/sd_054_scaffolded_onboarding_substrate_design.md; lever-B + SD-056 chain currently in validation provides the substrate prerequisite.
8. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21.

---

Status Snapshot (2026-05-30T01:10Z – nightly docs sync; SD-056 E2 action-conditional divergence preservation substrate landed today (lever B contrastive next-state InfoNCE auxiliary on world_forward per the 2026-05-28 substrate-design memo; resolves the V3-EXQ-571 root-cause finding cand_world_pairwise_dist=0.0; two new E2FastPredictor helpers; bit-identical OFF default; 539/539 contracts PASS; V3-EXQ-613 substrate-readiness PASS; V3-EXQ-569a behavioural falsifier hit a self-anchored-targets NaN and was superseded by V3-EXQ-569b with observation-anchored targets via /diagnose-errors); MECH-090 R-c continuation (nav_competence axis) pass 2 of 2 for commitment_closure:GAP-4 landed (CommitReadiness module wired at both BetaGate elevate sites; harness-push notify_outcome seam; V3-EXQ-592b grid extended to 4 arms for orthogonal-axis falsifier; 523/523 contracts PASS); MECH-341 parameter retune (entropy_lambda 0.05 -> 0.5; entropy_bias_scale 0.1 -> 1.0) per V3-EXQ-611c PASS interpretation routed to V3-EXQ-614 behavioural successor; runner FAIL/ERROR manifest-persistence fix retrofitted to match the PASS branch (4 new contract tests; 543/543 PASS); /governance cycle 2026-05-29 evening applied 6 failure-autopsies (598/606/596-602/603-followon-604-605/490g-cohort/MECH-341-cluster) + 13 evidence_quality_note appends + 7 pending_retest_after_substrate flags + 1 SD-033a narrow_supports_flag; CREATED substrate_queue entry scaffolded_sd054_onboarding (priority 1, unblocks 9 claims); Phase 3 writer auto-recovery + coordinator.db committed_at backfill landed (8/8 NULL rows resolved); GAP-4 Tier-1 library rebuild made cfg.use_dacc=True unconditional + switched C3_lift_vs_baseline metric to goal_norm_peak delta; central runner_status freshness ~5h at this read (768 cumulative, +7 since 2026-05-29T01:10Z; 183 PASS / 308 FAIL / 84 ERROR / 193 UNKNOWN); pending_review.md 1 -> 2 (V3-EXQ-483d FAIL pending autopsy + V3-EXQ-612b ERROR pending diagnose); queue 3 -> 1 (V3-EXQ-592c MECH-090 R-c validation on ree-cloud-3, supersedes 592b silent-drop))

• Substrate / governance landings since the 2026-05-29T01:10Z snapshot:
  • SD-056 E2 action-conditional divergence preservation (ree-v3 main 041a974; design doc REE_assembly/docs/architecture/sd_056_e2_action_conditional_divergence.md; plan-of-record memo REE_assembly/evidence/planning/e2_action_divergence_substrate_design.md; lit-pull REE_assembly/evidence/literature/targeted_review_e2_forward_model_action_divergence/SYNTHESIS.md) – substrate-level fix for the V3-EXQ-571 root-cause finding (2026-05-25). Adds InfoNCE-style auxiliary loss on E2.world_forward: positive (z_world_0, a_i) -> predicted z_world_1[i], negatives from K-1 in-batch sibling CEM candidates with different first-actions, temperature 0.1. Two new helpers on E2FastPredictor: cand_world_pairwise_dist (headline substrate-readiness diagnostic, methodological-gap publishable per SYNTHESIS verdict 3) and world_forward_contrastive_loss (returns unweighted CE; caller weights it via e2_action_contrastive_weight). Scope: applies to world_forward only, not predict_next_self (z_self is not the collapse site). Biology anchors: cerebellar internal model (Tanaka 2020) + prefrontal counterfactual rollout (Miyamoto 2023) + vestibular cerebellum corollary discharge (Cullen 2023). ML/AI anchors: Srivastava 2021 contrastive RSSM, Saanum 2024 PLSM failure diagnosis, InfoNCE foundation. Config knobs (E2Config + REEConfig.from_dims mirror): e2_action_contrastive_enabled (default False; bit-identical OFF), _weight (0.01), _temperature (0.1), _min_batch_classes (2). MECH-094 standard simulation_mode kwarg. 539/539 contracts + 7/7 preflight PASS with master OFF. V3-EXQ-613 substrate-readiness PASS; V3-EXQ-569a behavioural falsifier crashed twice with NaN in torch.multinomial (self-anchored InfoNCE targets caused E2 weights to diverge unboundedly); /diagnose-errors superseded 569a with V3-EXQ-569b using observation-anchored targets sampled from a rolling buffer of (z_world_0_pre_step, action_taken_one_hot, z_world_1_post_step_observed) triples captured each env tick.
  • MECH-090 R-c continuation (nav_competence axis) – pass 2 of 2 for commitment_closure:GAP-4 (design doc REE_assembly/docs/architecture/mech_090_commit_entry_predicate.md R-c continuation section). The 2026-05-28 landing covered the WITHIN-TICK DECISIVENESS axis (per-candidate score margin – Hanes & Schall 1996); this pass adds the ACROSS-TICK MOTOR-PROGRAM READINESS axis (Cisek & Kalaska 2010 affordance-preparation + Roesch / Calu / Schoenbaum 2007 dopaminergic readiness). New module ree_core/policy/commit_readiness.py (CommitReadiness + CommitReadinessConfig); pure-arithmetic, no nn.Module, no learned params. Maintains a [0,1] readiness EMA over per-tick outcome signals plus an explicit notify_outcome(value) harness-push seam; initial value 1.0 fail-open. Both R-c axes AND-compose at both BetaGate elevate sites (bistable + legacy). REEConfig knobs: use_mech090_readiness_conjunction (default False; auto-arms use_commit_readiness via OR-only resolver), mech090_readiness_floor (default 0.3 – V3-EXQ-592 seed 42 nav_competence=0.0 clearly fails to clear), commit_readiness_window (20), commit_readiness_ema_alpha (0.1; ~10-tick half-life), commit_readiness_initial (1.0). MECH-094 standard simulation_mode pattern. Per-tick outcome-signal source Phase 1: harness pushes via notify_outcome (committed_mode_curriculum.py wired); Phase 2 follow-on (separate session) wires env-emitted "mech090_readiness_outcome" key in agent.sense(). 523/523 contracts + 7/7 preflight PASS with both R-c master flags OFF, including 17 new MECH-090 R-c-nav-competence contracts. V3-EXQ-592b grid extended to 4 arms (ARM_0 baseline / ARM_2 GATED_NAV_COMP_ON / ARM_3 GATED_BOTH_ON / ARM_4 BOTH_GATES_OFF_HARNESS_FORCES_READY for orthogonal-axis falsifier discrimination per the design doc’s falsifier grid).
  • MECH-341 parameter retune (ree-v3 a45ca7f) – triggered by V3-EXQ-611c PASS interpretation grid (C1=True stratified fires, C2=False entropy bonus too small, C3=True diversity produced, R2c=True all ready). For 90% dominant class, bonus = lambda * freq = 0.045, far below observed score gaps 0.27-1.96; entropy_lambda raised 0.05 -> 0.5 (10x increase for C2 scale-commensurability); entropy_bias_scale raised 0.1 -> 1.0 for headroom. Config dataclass defaults + from_dims signature defaults + build_from_ree_config fallbacks all updated. 543/543 contracts PASS post-change (bit-identical when master OFF). V3-EXQ-611c interpretation routes to behavioural successor V3-EXQ-614 (B_only / ablate_B / ALL_ON per behavioral_diversity_isolation_plan.md R2.c rule).
  • Runner FAIL/ERROR manifest-persistence fix (ree-v3 41c3411) – experiment_runner.py FAIL branch (2267-2305) and ERROR branch (2229-2265) now invoke the three-call sequence (_result_manifest_exists + git_push_results + coordinator_client.report_result) that the PASS branch retrofitted on 2026-05-08 after the line-1394 UNKNOWN silent-drop bug. Root cause: V3-EXQ-592b on DLAPTOP-4 (2026-05-29T08:32:39Z) recorded FAIL + 4162s + a sentinel-written log line but the manifest was absent on disk, coordinator results table empty, no commit on origin/master. FAIL/ERROR with claimed-but-disk-missing manifest now WARN + release_active_claim + _pass_skip.add + continue (leaves queue entry) instead of silently removing. ERROR branch is gated on output_file non-empty so legitimate script crashes without a claimed manifest still proceed to queue removal. 4 new source-inspection contract tests in tests/contracts/test_runner_fail_branch_persists_result.py (C1 FAIL manifest-guard + skip-semantic; C2 FAIL ships manifest before queue-remove; C3 ERROR manifest-guard gated on output_file non-empty; C4 ERROR ships manifest before queue-remove). Preflight + contracts: 7/7 + 543/543 PASS.
  • /governance cycle 2026-05-29 evening – 6 user-confirmed failure-autopsies applied (598/606/596-602/603-followon-604-605/490g-cohort/MECH-341-cluster). Net claims.yaml edits: 13 evidence_quality_note appends (SD-033a + MECH-204 + 11 from 603-followon / 490g-cohort) + 7 pending_retest_after_substrate flags (Q-044/MECH-314/MECH-314a/Q-043/ARC-065 + MECH-313/Q-045 already had) + SD-033a narrow_supports_flag. Net substrate_queue.json: 1 CREATE (scaffolded_sd054_onboarding priority 1, unblocks 9 claims: Q-045/MECH-313/MECH-260/MECH-295/MECH-307/MECH-117/SD-049-Phase-2/ARC-030/Q-040; design_doc evidence/planning/sd_054_scaffolded_onboarding_substrate_design.md from sibling-session memo a425b9750a; 603c initial failure_record) + 1 amend ARC-046 (598 failure_record) + 1 amend MECH-341 (no-op, 611 failure_record already in place) + 1 amend scaffolded_sd054_onboarding cluster (603a + 603b + 604 + 605 failure_records). Net manifest edits: 2 reclassifications (596 + 602 non_contributory -> inconclusive_measurement). claims.json rebuilt 659 claims (byte-identical – evidence_quality_note doesn’t emit). pending_review.md cleared to 2 runner-only (V3-EXQ-483d FAIL pending autopsy + V3-EXQ-612b ERROR pending diagnose-errors).
  • Phase 3 writer auto-recovery + coordinator.db committed_at backfill (ree-v3 d3d3c7a + 61be94c + 21be280 + d3d3c7a) – sync_daemon.py _sync_to_origin gains opt-in PHASE3_AUTO_RESET_ON_REBASE_CONFLICT=1 self-heal path: on rebase failure with clean tree + writer-authored ahead commits, hard-resets to origin/<branch> and next tick re-materialises from DB/spool. Default off preserves the existing safe abort-and-wait. One-shot coordinator/backfill_committed_at.py helper walked results rows with committed_at IS NULL, checked origin/master for matching manifests (path lookup by queue_id prefix), and updated committed_at to manifest commit timestamps. Three landings progressively fixed candidate-search bugs (per-EXQ subdir walk; needle - vs _ normalisation; _partial/ subdir-name-gate removal). Hub run resolved 8/8 NULL rows against manifests on origin/master; post-write DB shows 0 NULL committed_at across 11 rows. Mac runner restart turned out unnecessary – serve.py _default_runner_extra_env already injects PHASE3_DISABLE_RUNNER_*_PUSH=1 when launching via /api/runner/v3/start (current Mac runner PID 96629 has all three env vars set per ps eww; residual legacy claim: / heartbeat DLAPTOP-4.local commits all dated to a pre-08:05Z earlier runner session).
  • GAP-4 Tier-1 library rebuild + V3-EXQ-483d / 490h queued (ree-v3 3eb2601) – per the 2026-05-29 V3-EXQ-490g-cohort autopsy Fork A routing. (1) experiments/_lib/goal_pipeline_tier1.py: cfg.use_dacc=True is now UNCONDITIONAL in build_config (moved out of the gap4_operating=True branch) so every cohort experiment gets dACC instantiation without per-script opt-in – closes the 483c primary diagnosis (agent.dacc=None across all 12 runs -> dacc_bias_nonzero_steps=0 unconditionally). (2) evaluate_tier1_cohort C3_lift_vs_baseline default metric switched from approach_commit_rate (saturated at 1.0 in OFF_OFF baseline under drive_floor=0.9 + goal_stream + reef – no headroom) to goal_norm_peak delta vs baseline (substrate-side, cross-claim-comparable; range 0.09-0.36 observed on 483c / 524a). New c3_lift_metric kwarg + _c3_lift_compare helper supports per-call override to override_signal_nonzero_steps (SD-037-specific) or legacy approach_commit_rate (debug only). (3) eval_tier1 records override_signal_nonzero_steps + new _override_signal_value helper. New scripts: V3-EXQ-490h (supersedes 490g; MECH-295 cascade Tier-1 retest, claim_ids=[MECH-295], priority 295) + V3-EXQ-483d (supersedes 483c; SD-037 broadcast override 4-arm with PAG/override_signal C2 + goal_norm_peak C3 + extra_config={use_dacc:True} explicit on every arm; claim_ids=[SD-037, MECH-280, MECH-281]; priority 290).
• Runner activity since the 2026-05-29T01:10Z snapshot: central evidence/experiments/runner_status.json reports 761 -> 768 cumulative completions (+7); all-time breakdown 183 PASS / 308 FAIL / 84 ERROR / 193 UNKNOWN (deltas PASS +5, FAIL +1, ERROR +1, UNKNOWN +0); last_updated 2026-05-28T17:26:40Z -> 2026-05-29T20:17:10Z – ~5h fresh at this read. Phase-3 coordinator -> central-index merge stayed caught up across the day. Recent tail dominated by the V3-EXQ-543g / 514j / 471a / 599 / 600 / 590a / 603b / 609 / 543h / 587 / 543j / 588b / 524a / 543k / 490g cluster previously dispositioned + V3-EXQ-611c PASS 2026-05-29T18:45Z (MECH-341 retune 6-arm validation, manifest-recovery re-run of V3-EXQ-611b) + V3-EXQ-612d PASS 2026-05-28T17:57Z (Phase 3 E2E smoke emit_outcome sentinel) + V3-EXQ-590a PASS 2026-05-25T08:40Z (EXQ-ISEF-004 rerun novelty bonus Goldilocks with checkpoint/resume).
• Queue depth: 3 -> 1 item – V3-EXQ-612 / 612c / 612d Phase-3 cutover smoke entries all drained (612d PASSed). New entry: V3-EXQ-592c (MECH-090 R-c commit-readiness gate validation on ree-cloud-3 with machine_affinity="ree-cloud-3" rather than "any" to deliver the validation on a different transport path; claimed 2026-05-29T22:46Z; estimated 180 min; supersedes V3-EXQ-592b which silent-dropped FAIL on DLAPTOP-4 2026-05-29T08:32Z and motivated the runner-side FAIL/ERROR-branch manifest-persistence fix). The MECH-341 retune validation V3-EXQ-611c PASSed earlier today and the MECH-090 R-c readiness validation V3-EXQ-592b was rerun as 592c after the silent-drop.
• Pending review: 1 -> 2 items (pending_review.md regenerated 2026-05-29T23:36Z; last review 2026-05-29T21:35Z). V3-EXQ-598b cleared via the evening governance cycle. New items: V3-EXQ-483d FAIL (Tier-1 library rebuild successor to 483c; C1+C2+C3+C4 PASS but C3_lift_vs_baseline FAIL 1/2 arms cleared; manifest tags SD-037=weakens; flagged for /failure-autopsy per user-confirmed routing; NOT added to reviewed_run_ids; stays pending) + V3-EXQ-612b ERROR (Phase 3 cutover cloud-2 smoke, 1.2s no-sentinel; stays pending for /diagnose-errors). V3-EXQ-611c + V3-EXQ-613 + V3-EXQ-544a all marked reviewed via discussed_experiment_dirs (substrate-readiness diagnostics, scoring_excluded).
• Bottleneck note (Layer-A E2 fix now LANDED; goal-pipeline substrate enrichment is the next bottleneck): SD-056 lever-B InfoNCE contrastive next-state landed today closes the upstream E2 world-forward per-candidate z_world collapse at the substrate level (V3-EXQ-613 PASS confirms the diagnostic substrate; V3-EXQ-569a -> 569b is testing whether the fixed substrate restores behavioural diversity downstream). The next-cycle bottleneck is the goal-pipeline / training-regime substrate enrichment identified by the 2026-05-29 V3-EXQ-490g cohort autopsy Cluster B disposition: the substrate-uniform monomodal-V_s monostrategy tail signature across 483c / 524a / 603 / 603a / 603b / 603c / 604 / 605 / 540a-e / 590a / 591 / 598 / 598b is the load-bearing constraint behind the diversity-cluster non_contributory chain. The new substrate_queue entry scaffolded_sd054_onboarding (priority 1, unblocks 9 claims: Q-045 / MECH-313 / MECH-260 / MECH-295 / MECH-307 / MECH-117 / SD-049-Phase-2 / ARC-030 / Q-040; design_doc evidence/planning/sd_054_scaffolded_onboarding_substrate_design.md) is the next substrate-implementation session-of-record. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding – not closed by either today’s autopsies.

Immediate Work Queue (This Cycle, 2026-05-30)

1. V3-EXQ-592c outcome watch – MECH-090 R-c commit-readiness gate validation on ree-cloud-3 (re-run of the silent-dropped 592b on DLAPTOP-4). ARM_0 GATED expecting total_committed_steps=0 + mech090_n_elevation_blocked >= 1; ARM_1 GATED_FORCED_READY expecting total_committed_steps > 0 + mech090_n_elevation_admitted >= 1. Joint PASS clears commitment_closure:GAP-4 partial -> done. Also tests end-to-end manifest shipping on FAIL post the runner-side FAIL/ERROR-branch fix landed today.
2. V3-EXQ-569b outcome watch – the SD-056 behavioural-validation falsifier on the fixed substrate (replaces V3-EXQ-569a after the self-anchored-targets NaN crash). Matched-entropy FP-2 falsifier per GAP-A R1.a/R1.b decision rule. PASS clarifies whether SD-056 substrate restores behavioural diversity downstream of the E2 forward-model fix.
3. /implement-substrate scaffolded_sd054_onboarding session-of-record – the next substrate-implementation session per the governance-applied 2026-05-29 V3-EXQ-490g-cohort autopsy Fork B + user-confirmed (A2) sub-lever. Design memo at evidence/planning/sd_054_scaffolded_onboarding_substrate_design.md specifies a three-phase P0/P1/P2 curriculum using SD-054 reef + bipartite-horizontal as a spatially-scaffolded start-state distribution (P0 agent spawns inside reef refuge band via a new env-side reef_bipartite_agent_spawn_in_reef_half kwarg, default False / bit-identical OFF), then anneals hazard_food_attraction 0.0 -> 0.7, proximity_harm_scale 0.05 -> 0.1, mech295_min_drive_to_fire 1.0 -> 0.01, mech307_conjunction_z_beta_threshold 0.6 -> 0.3 across P1, with P2 measurement on full target env. Commits to new scheduler module ree-v3/experiments/scaffolded_sd054_onboarding.py over extension of ARC-046 InfantCurriculumScheduler.
4. V3-EXQ-483d failure-autopsy – the Tier-1 library rebuild successor FAIL needs autopsy. C1+C2+C3+C4 PASS but C3_lift_vs_baseline FAIL 1/2 arms cleared; manifest tags SD-037=weakens. User-confirmed routing was to flag for autopsy without inline stamp.
5. V3-EXQ-612b /diagnose-errors – Phase 3 cutover cloud-2 smoke 1.2s no-sentinel ERROR. Resolve the missing sentinel + emit_outcome wiring before the next Phase 3 cutover smoke pass.
6. V3-EXQ-490h outcome watch – MECH-295 cascade Tier-1 retest with goal_norm_peak C3 metric (supersedes 490g) tests whether the GAP-4 library rebuild + unconditional use_dacc + dACC instantiation actually restores cross-arm divergence at the substrate level for MECH-295.
7. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21 drained-without-manifest BLOCK flagged by the governance verification gate; neither today’s autopsies nor governance closes it.
8. V3-EXQ-614 queueing decision – the MECH-341 retune behavioural successor (B_only / ablate_B / ALL_ON per behavioral_diversity_isolation_plan.md R2.c rule). Queue after V3-EXQ-611c sentinel routing confirms readiness; current substrate_queue MECH-341 status retune_implemented_pending_validation.
9. Phase 3 cutover go/no-go – writer auto-recovery + committed_at backfill landed today; pending the PHASE3_AUTO_RESET_ON_REBASE_CONFLICT=1 env-var activation in the hub writer’s systemd unit and the operational durability question.

---

Status Snapshot (2026-05-29T01:10Z – nightly docs sync; MECH-090 R-c commit-entry readiness conjunction substrate landed today (BetaGate.should_admit_elevation predicate; margin = sorted(scores)[1] - sorted(scores)[0], admits iff >= commit_readiness_floor=0.05; resolves commitment_closure_plan GAP-4 at substrate-readiness level; reading R-c synthesis-strongest per the 28-entry MECH-090 lit synthesis commit 9e68c5ca8a anchored on Cisek-Kalaska 2010 + Hanes-Schall 1996 + Roesch-Calu-Schoenbaum 2007; R-b Tandetnik 2021 retained as fallback; bit-identical OFF master; V3-EXQ-592b 2-arm GATED / GATED_FORCED_READY validation queued); MECH-341 retune (ree-v3 e02e77f) expanded stratified_select to BOTH committed and uncommitted branches in e3_selector.py resolving V3-EXQ-611 ARM_2 n_stratified_fired=0 zero-fires failure (committed branch was never entered during validation); V3-EXQ-611b 6-arm factorial parameter sweep queued and runner-claimed DLAPTOP-4.local 2026-05-28T17:26Z; E2 action-conditional divergence lit-pull SYNTHESIS (6 entries balanced across ML world-model + biology forward-model, lit_conf 0.78) verdict option (ii) fix E2 as architecturally faithful target with lever B (InfoNCE contrastive next-state) chosen over PLSM / SWIRL in companion substrate-design memo (REE_assembly 7cb1200332); coord-env Mac runner-start shadow-env injection fix + heartbeat coordination_mode field + /queue-experiment skill verification step landed (REE_assembly fc08812b62 + ree-v3 9fc0e02); central runner_status freshness ~7h45m at this read (761 cumulative, +5 since 2026-05-28T01:10Z; 178 PASS / 307 FAIL / 83 ERROR / 193 UNKNOWN); cross-machine unchanged; pending_review.md 1 -> 1 (V3-EXQ-598b FAIL unchanged); queue 1 -> 3 (V3-EXQ-610 drained; 3 Phase-3 cutover smoke entries V3-EXQ-612 / 612c / 612d in flight))

• Substrate landings since the 2026-05-28T01:10Z snapshot (two major substrate claims + infra):
  • MECH-090 R-c commit-entry readiness conjunction (per ree-v3 CLAUDE.md MECH-090 section, landed 2026-05-28 in the active implement-substrate-mech090-rc-conjunction session; ree-v3 + REE_assembly architectural design doc REE_assembly/docs/architecture/mech_090_commit_entry_predicate.md) – BetaGate.should_admit_elevation predicate added with margin = sorted(scores)[1] - sorted(scores)[0]; admits commitment iff margin >= commit_readiness_floor (default 0.05). Wraps both beta_gate.elevate() call sites in REEAgent.select_action (bistable branch + legacy branch). Resolves commitment_closure_plan.md GAP-4 at the substrate-readiness level after V3-EXQ-592 seed 42 (2026-05-21) showed the legacy rv-only commit-entry predicate is satisfiable by degenerate trivial-predictability (rv=2.7e-5 with nav_competence=0.0). Reading R-c single-gate conjunction (synthesis-strongest) per REE_assembly/evidence/literature/targeted_review_connectome_mech_090/synthesis.md (commit 9e68c5ca8a, 28 entries) – anchored on Cisek & Kalaska 2010 (affordance-competition), Hanes & Schall 1996 (FEF accumulator-to-threshold), Roesch / Calu / Schoenbaum 2007 (dopaminergic readiness signal). R-a (rv-only correct) not defensible post-pass; R-b (rv-only entry + downstream propagation gate, Tandetnik 2021) retained as fallback if validation fails. Knobs on HeartbeatConfig (NOT surfaced through REEConfig.from_dims to avoid concurrent-session signature conflict with the MECH-341 retune): use_commit_readiness_gate (default False; bit-identical OFF master), commit_readiness_floor (0.05), commit_readiness_strict_single_candidate (False; permissive single-candidate handling). Per-episode diagnostics on BetaGate.get_state: mech090_n_elevation_admitted / _blocked / _single_candidate / _last_readiness_score_margin. MECH-094 N/A (control-state-transition predicate; no simulation-write surface). 506/506 contracts PASS with master OFF + 7 unit tests on the BetaGate primitive PASS. V3-EXQ-592b 2-arm validation queued: ARM_0 GATED at floor=0.05 expecting total_committed_steps=0 + mech090_n_elevation_blocked >= 1 + running_variance below commitment_threshold (the gate is the load-bearing block); ARM_1 GATED_FORCED_READY with experiment-side score_bias injection forcing margin >= 0.10 expecting total_committed_steps > 0 + mech090_n_elevation_admitted >= 1 (gate does not permanently lock out commitment when readiness clears). Joint PASS clears commitment_closure:GAP-4 partial -> done.
  • MECH-341 retune (ree-v3 e02e77f) – stratified_select call-site expanded from committed-only to BOTH committed and uncommitted branches in ree_core/predictors/e3_selector.py lines 848-870. Resolves V3-EXQ-611 ARM_2 (OPT2 stratified_select only) n_stratified_fired=0 across all 3 seeds: the committed branch was never entered in the validation episodes (running_variance never fell below effective_threshold), so OPT2 had no opportunity to fire. The 2026-05-28 patch invokes stratified_select in BOTH branches when the substrate is enabled AND the pool admits >=2 first-action classes; bit-identical when score_diversity is None or sub-flag is False. MECH-094 preserved by simulation_mode=False kwarg at both call sites. Architectural rationale: the substrate_queue’s failure-record language (“substrate-natural pool diversity gets preserved through softmax-sample-across-class-representatives”) describes a categorical-preservation semantic that applies regardless of commit state. V3-EXQ-611b 6-arm factorial validation queued (3 option groups OPT1_only / OPT2_only / BOTH x 2 entropy_bias_scale values 1.0 / 2.0; ALL_OFF baseline anchored to V3-EXQ-611 ARM_0_ALL_OFF manifest already on origin/master; acceptance C1 primary = n_stratified_fired > 0 across all OPT2/BOTH seeds + (C2 bonus scale-commensurate OR C3 selected-class diversity preserved on majority of seeds); runner-claimed DLAPTOP-4.local 2026-05-28T17:26Z; estimated 200 min; experiment_purpose=diagnostic – substrate-readiness retunes do NOT weight claim confidence per Phase-3 governance rules).
  • E2 action-conditional divergence substrate-design memo (REE_assembly 7cb1200332, evidence/planning/e2_action_divergence_substrate_design.md) – picks lever B (contrastive next-state, Srivastava et al. 2021 style) over A/PLSM and C/SWIRL per the 2026-05-28 6-entry lit-pull SYNTHESIS verdict (REE_assembly 04bc1f3727; balanced across ML world-model + biology forward-model literatures, lit_conf 0.78). Specifies the InfoNCE-style auxiliary on E2.world_forward with K-1 in-batch negatives drawn from sibling CEM candidates with different first-actions; four config knobs (e2_action_contrastive_enabled default False bit-identical OFF, _weight 0.01, _temperature 0.1, _min_batch_classes 2); MECH-094 simulation_mode kwarg falling through to tensor(0.0); cand_world_pairwise_dist diagnostic helper at e2_fast.py named as a publishable standalone metric per the SYNTHESIS methodological-gap finding (no published paper reports per-action pairwise distance between predicted latents at one-step horizons as headline metric); UC1-UC5 acceptance criteria for V3-EXQ-NEW-1 substrate-readiness diagnostic with UC3 (direction of change in cand_world_pairwise_dist) as the load-bearing PASS condition; six-step plan-of-record sequencing across separate sessions (memo -> /implement-substrate -> /queue-experiment -> runner -> /queue-experiment -> runner -> governance); V3-EXQ-569a as the matched-entropy FP-2 falsifier successor applying GAP-A R1.a/R1.b decision rule on the fixed substrate. Three honest disclaimers: (1) the contrastive task being learnable does not by itself imply behavioural diversity emerges; (2) PLSM/SWIRL are valid fallbacks if lever B fails; (3) option (i) GAP-B one-hot bypass is still a tactical alternative.
• Infra activity since the 2026-05-28T01:10Z snapshot:
  • coord-env Mac runner-start fix (REE_assembly fc08812b62 + ree-v3 9fc0e02) – three-part fix for the user-reported issue that the Mac runner started via explorer /api/runner/v3/start button has no COORDINATION_MODE / COORDINATOR_URL / COORDINATOR_TOKEN, so coordinator_client._ENABLED=False and report_result is a no-op. (1) REE_assembly/serve.py start_runner(): default-inject shadow env from coordinator.env when extra_env is None and env file is configured; bit-identical for existing shadow/coordinator panel callers that pass explicit extra_env. (2) ree-v3/runner_remote_control.py write_heartbeat(): surface coordination_mode field auto-read from os.environ so cross-machine status is visible without an SSH audit; backward-compatible additive field. (3) /queue-experiment skill (both .claude/ and .agents/ mirrors per CLAUDE.md mirror rule): add a verification step that scans runner_heartbeats/*.json and warns if any active runner is in git mode or missing the field. (4) cloud_workers.md (gitignored local-only operator notes): document the heartbeat verification recipe so cloud workers can be spot-checked without SSH. Pre-flight smoke: serve.py + runner_remote_control.py ast-parse OK; write_heartbeat default mode=’git’ verified; with COORDINATION_MODE=shadow exported, write_heartbeat surfaces ‘shadow’ correctly; _default_runner_extra_env() returns populated env dict with mode=shadow + url=http://10.8.0.1:8787 + token=43 chars from this Mac’s coordinator.env.
  • IGW housekeeping batch (4 plan-doc syncs + 1 workset regen, all governance-side; no claim landings): IGW-008 GAP-A plan resync (behavioral_diversity_isolation_plan.md row 1 status partial -> blocked_pending_substrate; resume_condition rewritten against the 2026-05-25 V3-EXQ-571 root-cause finding); IGW-010 GAP-C plan-doc refresh + inter_governance_workset regen (row 3 in_progress -> blocked_pending_substrate per V3-EXQ-603c FAIL non_contributory + V3-EXQ-611 FAIL cluster-absorbed into the V3-EXQ-591 autopsy; gov-correction substrate_ceiling V3 routing); IGW-011 GAP-D doc-sync + R4.b flag (row 4 in_progress -> pending_governance_stamp against the landed manifests V3-EXQ-550 FAIL supports MECH-269 + V3-EXQ-601 PASS supports MECH-269b); MECH-341 retune session close updated row 2 GAP-B owner_exq + resume_condition against the landed retune.
• Runner activity since the 2026-05-28T01:10Z snapshot: central evidence/experiments/runner_status.json reports 756 -> 761 cumulative completions (+5); all-time breakdown 178 PASS / 307 FAIL / 83 ERROR / 193 UNKNOWN (deltas PASS +0, FAIL +3, ERROR +2, UNKNOWN +0); last_updated 2026-05-27T18:05:04Z -> 2026-05-28T17:26:40Z – ~7h45m fresh at this read. The Phase-2 coordinator -> central-index merge remains caught up from yesterday’s improvement. Last 30 entries are dominated by the V3-EXQ-543g / 514j / 471a / 599 / 600 / 590a / 603b / 609 / 543h / 587 / 543j / 588b / 524a / 543k / 490g cluster previously dispositioned. The MECH-090 R-c readiness conjunction validation V3-EXQ-592b and the MECH-341 retune validation V3-EXQ-611b queued by today’s substrate-landing sessions are runner-claimed (611b on DLAPTOP-4.local @17:26:40Z per the MECH-341 retune session close note; 592b similarly claimed by the active MECH-090 R-c session) but their manifests are not yet in the central index at this read.
• Queue depth: 1 -> 3 items, all Phase-3 cutover smoke – V3-EXQ-610 (yesterday’s INV-074 retest-after-substrate, priority 28, machine_affinity=any, estimated_minutes=180) drained (presumably claimed by a runner; manifest landing pending the next read). New entries: V3-EXQ-612 (Phase 3 cutover end-to-end smoke, DLAPTOP-4.local claimed 2026-05-28T17:24:31Z, priority 1, estimated 1 min; trivial smoke driving the claim->result->writer->queue-writer Phase 3 cycle); V3-EXQ-612c (ree-cloud-2 pending; supersedes V3-EXQ-612b which ran fine on cloud-2 but the script lacked the verdict: PASS stdout sentinel so the runner classified ERROR and skipped report_result; script now prints verdict: PASS); V3-EXQ-612d (ree-cloud-3 pending; supersedes V3-EXQ-612c manifest path was empty in runner because the script did not call experiment_protocol.emit_outcome; now wired; single try then stop chasing the E2E demo if this fails). The V3-EXQ-611b MECH-341 retune validation and V3-EXQ-592b MECH-090 R-c readiness validation that the 2026-05-28 substrate-landing sessions queued do NOT appear in items[] at this read – both were runner-claimed earlier in the day per the respective session close notes.
• Pending review: 1 -> 1 item unchanged (pending_review.md regenerated 2026-05-27T17:40:25Z; last review 2026-05-27T17:35Z) – V3-EXQ-598b (commitment_closure GAP-1 SD-033a bias-head trainable ablation, claim_ids=[MECH-262, SD-033a]) FAIL 2026-05-27T12:03Z evidence_direction=does_not_support. No fresh failure-autopsy yet against 598b; queued for the next governance pending walk.
• Bottleneck note (substrate-design phase for E2 fix landed today; Layer-A cause still unaddressed in code): The upstream E2 world-forward per-candidate z_world collapse identified in the 2026-05-25 V3-EXQ-571 root-cause investigation remains the structural root cause of the score_bias-chain flatness; today’s E2 action-divergence design memo makes lever B (InfoNCE contrastive next-state) the plan-of-record fix per the 2026-05-28 lit-pull SYNTHESIS verdict, but landing it in code is deferred to a separate /implement-substrate session and is the next session-of-record after V3-EXQ-611b and V3-EXQ-592b return validation signal. Today’s MECH-341 retune and MECH-090 R-c readiness conjunction address the Layer-B (E3 scoring) and the commit-entry predicate respectively, but neither removes the Layer-A (E2 forward-model collapse) cause. The 2026-05-28 work consolidates two substrate-readiness landings + the lever-B design plan-of-record decision; the heavy code lift (lever B implementation) is the next-cycle target. The V3-EXQ-543k disposition gap carried forward from 2026-05-21 remains outstanding – not closed today.

Immediate Work Queue (This Cycle, 2026-05-29)

1. V3-EXQ-611b outcome watch – the MECH-341 retune 6-arm parameter sweep is runner-claimed DLAPTOP-4.local 2026-05-28T17:26:40Z and is the load-bearing acceptance test for the call-site expansion (C1 primary). PASS routes governance walk to behavioural successor + R2.c rule; FAIL with C1=false routes to /diagnose-errors on e3_selector wiring; FAIL with C1=true and C2/C3=false routes to algorithm-level Option-2 substrate revisit.
2. V3-EXQ-592b outcome watch – the MECH-090 R-c readiness conjunction 2-arm validation should produce ARM_0 GATED (total_committed_steps=0 + mech090_n_elevation_blocked >= 1) AND ARM_1 GATED_FORCED_READY (total_committed_steps > 0 + mech090_n_elevation_admitted >= 1). Joint PASS clears commitment_closure:GAP-4 partial -> done. If ARM_0 FAILs (gate does not block under expected conditions) the R-c reading is falsified and routing reverts to R-b fallback (Tandetnik 2021 staged conservative).
3. E2 lever-B /implement-substrate session-of-record – the next plan-of-record session per the 2026-05-28 E2 action-divergence substrate-design memo. Lands the InfoNCE-style auxiliary contrastive loss on E2.world_forward with K-1 in-batch negatives drawn from sibling CEM candidates with different first-actions; four config knobs (e2_action_contrastive_enabled default False bit-identical OFF, _weight 0.01, _temperature 0.1, _min_batch_classes 2); MECH-094 simulation_mode kwarg falling through to tensor(0.0); cand_world_pairwise_dist diagnostic helper at e2_fast.py. Substrate-readiness V3-EXQ-NEW-1 + matched-entropy V3-EXQ-569a falsifier successor follow.
4. V3-EXQ-610 INV-074 manifest watch – the IGW-027 retest-after-substrate INV-074 / Layer-B crystallization necessity run (queued yesterday, priority 28, machine_affinity=any, estimated 180 min) drained from items[] today; manifest landing pending. Outcome routes the INV-074 / MECH-334 cluster directly under the new Layer-B substrate.
5. V3-EXQ-598b governance walk – carried forward from yesterday. The FAIL pending review (commitment_closure GAP-1 SD-033a bias-head trainable ablation, evidence_direction=does_not_support) needs the next governance cycle to walk and either (a) governance-stamp non_contributory if the trainable-vs-frozen comparison cannot fire due to upstream substrate (consistent with V3-EXQ-571 z_world collapse), (b) governance-stamp weakens SD-033a if the comparison did fire cleanly, or (c) route to /failure-autopsy if test fairness is ambiguous.
6. ARC-046 InfantCurriculumScheduler Phase-0 exit-gate substrate – carried forward. /implement-substrate target queued via substrate_queue per the V3-EXQ-591 failure autopsy (Phase 0->1 advancement signal tune H_POS_FRAC_OF_MAX threshold OR replace with z_goal-norm / residue-progression gate). Required to unblock the infant_substrate GAP-14 + curriculum-vs-flat ARC-046 comparison.
7. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21 drained-without-manifest BLOCK flagged by the governance verification gate.
8. Phase-3 cutover smoke V3-EXQ-612 / 612c / 612d outcome watch – the three queued E2E demo entries are the final piece of operational validation before the PHASE3_GIT_WRITER_READY=True flag flip; cloud-2 / cloud-3 trips test the runner_remote_control + coordinator wiring under the post-fix heartbeat field. The 612d note “Single try, then stop chasing the E2E demo if this fails” is the operational stop-condition.
9. Fleet contraction triage (carry-forward) – only DLAPTOP-4.local has been writing within the last 24h since the 2026-05-27 snapshot. Decide whether to restart cloud workers or accept a Mac-only fleet through the upcoming /implement-substrate lever-B session.

---

Status Snapshot (2026-05-28T01:10Z – nightly docs sync; MECH-341 e3_score_diversity substrate landed today as Layer-B post-CEM diversity-preservation regulator with togglable entropy-bonus + stratified-select sub-flavours; governance cycle 2026-05-27 dispositions: V3-EXQ-543l per-claim 4-split (ARC-062 weakens narrow_supports_flag / MECH-309 supports first trained-policy entry / INV-074 + MECH-334 non_contributory) + V3-EXQ-591 cluster-absorbed into 591 autopsy + Q-045 / MECH-313 / MECH-260 routed substrate_ceiling V3 (NOT substrate_conditional V4 per user-flagged correction); substrate_queue extended with ARC-046 InfantCurriculumScheduler exit-gate + MECH-341 retune entries; closure-drift /governance lint script + brain-map prefix mapping + 4 new regions + validator tightening landed; central runner_status freshness recovered to ~7h after the prior ~76h stall; cross-machine 1540 -> 1543; pending_review.md 0 -> 1 (V3-EXQ-598b FAIL); queue 0 -> 1 (V3-EXQ-610 INV-074 retest))

• Substrate landings since the 2026-05-27T01:10Z snapshot (single major claim landing):
  • MECH-341 e3_scoring_preserves_trajectory_class_diversity (ree-v3 547faa3 + REE_assembly 7cbb13833b) – new ree_core/predictors/e3_score_diversity.py module (E3ScoreDiversity + E3ScoreDiversityConfig + E3ScoreDiversityDiagnostics + build_from_ree_config). Layer-B (post-CEM scoring) diversity-preservation substrate triggered by V3-EXQ-608 P2 (2026-05-26T02:58Z) majority R2a_e3_collapse_confirmed_large_gap: with SP-CEM main-path delivering frac_pre_ge2=1.0, E3 scoring still collapsed to a single class with mean_top2_class_gap 0.27-0.60 (LARGE-gap; rules out option 3 jittered tie-breaking; routes to options 1 + 2 per behavioral_diversity_isolation_plan.md). Two togglable sub-flavours under one master, MECH-314a/b/c-style: Option 1 entropy_bonus (per-candidate positive bias proportional to first-action class frequency, composed AFTER the dACC / lateral_pfc / ofc / mech295 / curiosity / tonic_vigor score_bias chain and BEFORE last_scores / softmax) + Option 2 stratified_select (partition by first-action class, argmin within class, softmax-sample across class-representatives; replaces argmin in the committed-path selection at e3_selector.py:811-820). Pure-arithmetic regulator (no nn.Module inheritance, no learned parameters); sibling pattern to MECH-313 NoiseFloor + MECH-314 StructuredCuriosity + MECH-320 TonicVigor. REEConfig.use_e3_score_diversity master + 6 sub-knobs (use_e3_diversity_entropy_bonus, use_e3_diversity_stratified_select, e3_diversity_entropy_lambda 0.05, e3_diversity_entropy_bias_scale 0.1, e3_diversity_stratified_temperature 1.0, e3_diversity_min_classes_for_stratification 2) all default bit-identical OFF. 506/506 contracts + 7/7 preflight PASS with master OFF (regression-clean 2026-05-27). MECH-094: both methods accept simulation_mode argument; when True, apply_entropy_bonus returns zeros[K] and stratified_select returns None (caller falls through to legacy argmin); inline gates are defensive (current wired call site E3Selector.select is invoked only from waking REEAgent.select_action paths). Design doc REE_assembly/docs/architecture/mech_341_e3_score_diversity_preservation.md (full Status / Solution / Architecture-Context / Falsifiability / ML-AI-engineering-notes / Plan-of-record references). claims.yaml MECH-341 implementation_note + updated evidence_quality_note (v3_pending retained pending R2.c rule via V3-EXQ-611 PASS). behavioral_diversity_isolation_plan.md status-table row 2 + GAP-B node frontmatter updated to IMPLEMENTED + Section 332 supersedes notice. Validation V3-EXQ-611 queued priority 260 (4-arm ALL_OFF / OPT1_ONLY / OPT2_ONLY / BOTH_ON substrate-readiness diagnostic on the EXQ-608 env + metric stack; experiment_purpose=diagnostic; estimated_minutes=150; renumbered from 610 after collision with INV-074 already at that ID). Behavioural validation (Phase P3 B_only / ablate_B / ALL_ON arms per the isolation plan + R2.c rule) deferred to a successor queued after EXQ-611 PASS. User-decision-via-AskUserQuestion: “Both options, togglable” per the implement-substrate Step 3 design-plan presentation (rationale: matches MECH-314a/b/c precedent + lets Q-054 falsifier dissociate empirically).
• Governance / substrate-queue / closure-drift activity since 2026-05-27T01:10Z (4 governance-side commits + 1 brain-map commit + 1 closure-drift lint commit):
  • Governance cycle 2026-05-27 (REE_assembly 4856a3dcdb) – Step 0 cleared 3 stale IGW auto-spawn claims (IGW-026 6.5h, IGW-036 12.5h, IGW-031 23.2h). Pending walk 5 -> 1 indexed (V3-EXQ-598b correctly held for the next walk after today’s FAIL). Decisions applied: (a) V3-EXQ-543l per-claim 4-split governance-stamped (ARC-062 weakens narrow_supports_flag / MECH-309 supports first trained-policy entry / INV-074 + MECH-334 non_contributory missing-prerequisite); (b) V3-EXQ-591 manifest evidence_direction does_not_support -> non_contributory cluster-uniform 4th member; ARC-046 NOT weakened; (c) V3-EXQ-603c cluster-absorbed into the 591 autopsy; Q-045 / MECH-313 / MECH-260 initially routed epistemic_category=substrate_conditional pending V4 substrate (EXQ-445h remains valid MECH-260 support); (d) V3-EXQ-598b flagged for /failure-autopsy at the time (revisited after today’s FAIL completion); (e) V3-EXQ-611 diagnostic non_contributory by design; reviewed; retune routed via substrate_queue priority-2. Step 5b closure-drift: bumped arc_062 GAP-B + infant_substrate GAP-14 (status in-progress -> blocked); added closure_plan.last_updated frontmatter on 2 of 8 plans. drifted_nodes=13 unchanged; plans_missing_last_updated 8 -> 6. substrate_queue extended ARC-062 failure_record + added ARC-046 (InfantCurriculumScheduler exit-gate fix) + MECH-341 (entropy_bonus_scale retune + stratified trigger condition revision) entries. claims.json rebuilt 658 claims. validate_claims OK strict.
  • Governance correction 2026-05-27 (REE_assembly ac56ba507b) – user-flagged category error from the same-day governance stamping. Routing Q-045 / MECH-313 / MECH-260 to substrate_conditional pending V4 was a scope-deferral disguised as a category: their test bed (SD-054 reef + ARC-062 gated-policy + 4-arm ablation) is V3-scoped by design, and the 591 autopsy cluster reading is substrate_enrichment-within-V3, not a V4 prerequisite. Correction landed: epistemic_category substrate_conditional -> substrate_ceiling on Q-045 + MECH-313 + MECH-260; evidence_quality_note rewritten on all 3 to drop V4 framing; ARC-046 prereq-(b) note rewritten in substrate_queue; substrate_queue ARC-046 ready_blocked_by + depends_on_unresolved fixed.
  • Closure-drift /governance step + lint script (REE_assembly 01e5f79e7d) – new scripts/check_closure_drift.py (warn-only lint detecting closure_plan nodes whose owner_exq has reached a terminal state but whose status is still in-progress/blocked/upstream-blocked/partial, plus missing closure_plan.last_updated); wired into scripts/governance.sh as Step 3c; .claude/skills/governance/SKILL.md + .agents/skills/governance/SKILL.md updated with the closure-map reconciliation step that consumes the drift report. Initial drift report evidence/planning/closure_drift.md covers 12 drifted nodes across 8 plans (arc_062 GAP-B + GAP-D; commitment_closure GAP-4; infant_substrate GAP-11/12/13/14; self_attribution GAP-1; sleep_substrate GAP-2; behavioral_diversity_isolation GAP-A/B/D).
  • Brain-map prefix mapping + new regions + validator tightening (REE_assembly 4039a0dbaa) – 13 mechanical prefix mappings (acetylcholine/noradrenaline/orexin/gabaergic -> neuromodulation; bg -> basal_ganglia; commit_gate/hypothesis_tag/broadcast/theta_gamma -> control_plane; salience -> cingulate; lpb -> harm_stream; homeostatic_accumulation -> harm_stream; cortical_world_model -> e1); 4 new regions (TPJ, motor, peripersonal_space, visual_streams) with SVG placeholders; scope normalization (partial -> in_scope for now-expressed regions); validate_brain_region_map.py tightened to WARN on any claim prefix not in known_anatomy_prefixes AND not in explicit non_anatomy_prefixes allowlist; validator wired into scripts/governance.sh as warn-only Step 3d between Step 3c closure-drift and Step 4 claims.json rebuild; brain_map.html version stamp bumped. Brain map now has 20 regions + 7 engineering nodes (was 17+7); 0 unmapped_prefixes per live /api/brain-map.
  • Closure-frontmatter sync (behavioral_diversity_isolation_plan) (REE_assembly cf3012570a) – retrofit closure_plan YAML frontmatter onto evidence/planning/behavioral_diversity_isolation_plan.md so the Explorer closure card stops rendering it as frontmatter_pending placeholder. 8 nodes: 4 primary (Theories 1-4 = Layer A/B/C/D = GAP-A..GAP-D) + 4 secondary (Theories 5-8 = GAP-E..GAP-H, status=deferred). Closure API verified: card now renders frontmatter_pending=False with weighted progress 0.425.
  • Failure-autopsy artifacts: V3-EXQ-543l (REE_assembly 72bab05c93) ARC-062/MECH-309/INV-074/MECH-334 GAP-B escalated falsifier FAIL branch-e – substrate-enrichment-first routing recommended (GAP-C/D substrate-landed 2026-05-17 supplies the rule_state routing consumer; V3-EXQ-598b is the discriminating experiment); V3-EXQ-591 (REE_assembly cfedfd1353) ARC-046 curriculum-vs-flat 3-arm x 5-seed substrate-uniform FAIL (1/7 criteria, only C3 trivially saturating; curriculum scheduler stuck Phase 0 because H_pos advancement gate ~3.48 is structurally unreachable under random-policy 2000-episode training; z_goal collapses to ~1e-7 in every arm; C5/C6/C7 emit -1.0 sentinels in every seed; substrate-uniform z_goal-zero cluster pattern with V3-EXQ-540 / 603 / 590a / 591; ARC-046 untouched / non_contributory / pending_retest_after_substrate; /implement-substrate target: InfantCurriculumScheduler Phase 0->1 advancement signal tune H_POS_FRAC_OF_MAX threshold OR replace with z_goal-norm / residue-progression gate).
  • Stale-text follow-on cleanup: arc_062_rule_apprehension_plan.md GAP-C + GAP-D substrate_notes + commitment_closure_plan.md GAP-1 resume_condition rewritten to permissive-gate language matching the autopsy’s substrate-enrichment-first routing; inter_governance_workset.v1.json regenerated so IGW-20260527-012 agent_brief picks up the new resume_condition text (REE_assembly 70c991836d).
  • Phase3 writer review followups landed (ree-v3 86430b7 + a0b9da1..86430b7) – MED-A use _PHASE3_COMMIT_PREFIX constant in commit_msg construction; MED-B fsync parent dir after os.replace in atomic working-tree write; LOW-A author StaleOriginRefTest seed with phase3: prefix so test exercises the non-FF rejection path; LOW-C apply _validate_batch_size-style hardening to PHASE3_QUEUE_FILE/PHASE3_ASSEMBLY_BRANCH/SYNC_INTERVAL env knobs. Suite: 85/85 PASS. All review followups now landed; cluster is ready for the PHASE3_GIT_WRITER_READY=True flag flip pending operational durability question.
• Runner activity since the 2026-05-27T01:10Z snapshot: central evidence/experiments/runner_status.json reports 753 -> 756 cumulative completions (+3); all-time breakdown 178 PASS / 304 FAIL / 81 ERROR / 193 UNKNOWN; last_updated 2026-05-23T21:06:24Z -> 2026-05-27T18:05:04Z – the central file freshness has recovered to ~7h at this read (was ~76h yesterday), so the Phase-2 coordinator -> central-index merge caught up after the 2026-05-27 governance cycle staging. The +3 completions resolve to the V3-EXQ-543j / 588b / 524a / 543k / 490g cluster that the governance pending-walk dispositioned this morning. Cross-machine aggregate from per-host files: DLAPTOP-4.local 595 (+3; latest V3-EXQ-611 2026-05-27T13:02Z) + ree-cloud-1 243 + ree-cloud-2 184 + ree-cloud-3 142 + ree-cloud-4 141 + ree-worker-3 133 + EWIN-PC 77 + Daniel-PC 28 = 1543 cumulative across hosts (+3 vs yesterday’s 1540). Only DLAPTOP-4.local wrote within the last 24h; ree-cloud-3 is at ~42h stale (was ~18h yesterday); the other six hosts remain stale since 2026-05-21 or earlier. Fleet contraction continues.
• Queue depth: 0 -> 1 item – V3-EXQ-610 v3_exq_610_inv074_crystallization_necessity.py (INV-074, priority 28, machine_affinity=any, estimated_minutes=180) is the IGW-20260527-027 INV-074 retest-after-substrate with MECH-341 flags active per the 2026-05-27 IGW-027 close (commit a394dae; use_e3_score_diversity=True + entropy_bonus enabled). V3-EXQ-611 (the MECH-341 4-arm substrate-readiness diagnostic queued priority 260 in the MECH-341 session close note) does NOT appear in items[] at this read despite priority 260 being above V3-EXQ-610’s priority 28 – worth a manual recheck whether 611 was drained or was never persisted in the queue file.
• Pending review: 0 -> 1 item (pending_review.md regenerated 2026-05-27T17:40:25Z; last review 2026-05-27T17:35Z) – V3-EXQ-598b (commitment_closure GAP-1 SD-033a bias-head trainable ablation, claim_ids=[MECH-262, SD-033a]) FAIL 2026-05-27T12:03Z evidence_direction=does_not_support. The runner finished the ARM_0 frozen / ARM_1 trainable comparison and the permissive gate (manifest exists + outcome in {PASS, FAIL}) holds, so the disposition routes through the next governance pending walk rather than via /failure-autopsy.
• Bottleneck note (Layer-B substrate added; Layer-A cause unchanged): The upstream E2 world-forward per-candidate z_world collapse identified in the 2026-05-25 V3-EXQ-571 root-cause investigation remains the structural root cause of the score_bias-chain flatness; today’s MECH-341 substrate landing addresses the Layer-B (E3 scoring) symptom but does not remove the Layer-A (E2 forward-model collapse) cause. The 2026-05-27 governance correction (Q-045 / MECH-313 / MECH-260 substrate_conditional V4 -> substrate_ceiling V3) is the most consequential governance act of the cycle: it keeps the diversity cluster as a V3-scoped substrate-enrichment problem rather than punting to V4. ARC-062 / MECH-309 picked up its first contributory trained-policy entry today (V3-EXQ-543l per-claim split: MECH-309 supports first trained-policy entry; ARC-062 weakens narrow_supports_flag), and V3-EXQ-598b’s FAIL pending review will resolve whether the next-action substrate-enrichment pass is the ARC-046 InfantCurriculumScheduler Phase-0 exit gate or the MECH-341 entropy retune. The V3-EXQ-543k disposition gap carried forward from the 2026-05-21 drained-without-manifest BLOCK remains outstanding – not closed by either failure-autopsy today.

Immediate Work Queue (This Cycle, 2026-05-28)

1. V3-EXQ-611 queue-state recheck – the MECH-341 4-arm substrate-readiness diagnostic was queued at priority 260 in the MECH-341 session close note but is absent from experiment_queue.json at this read. Confirm whether it was drained (manifest written / runner picked it up) or whether the queue write was lost; re-queue if necessary so the MECH-341 substrate-readiness gate (PASS = either single-option arm produces selected_action_classes_count >= 2 with frac_pre_ge2 >= 0.5) can fire before the behavioural-validation successor.
2. V3-EXQ-598b governance walk – the FAIL pending review (commitment_closure GAP-1 SD-033a bias-head trainable ablation, evidence_direction=does_not_support) needs the next governance cycle to walk and either (a) governance-stamp non_contributory if the trainable-vs-frozen comparison cannot fire due to upstream substrate (consistent with V3-EXQ-571 z_world collapse), (b) governance-stamp weakens SD-033a if the comparison did fire cleanly, or (c) route to /failure-autopsy if the test fairness is ambiguous.
3. V3-EXQ-610 INV-074 outcome watch – the IGW-027 retest-after-substrate run with MECH-341 entropy_bonus active is the first V3-EXQ-610 invocation that tests INV-074 (Layer-B crystallization necessity) under the new Layer-B substrate. Outcome routes the INV-074 / MECH-334 cluster directly.
4. ARC-046 InfantCurriculumScheduler Phase-0 exit-gate substrate – /implement-substrate target queued via substrate_queue per the V3-EXQ-591 failure autopsy (Phase 0->1 advancement signal tune H_POS_FRAC_OF_MAX threshold OR replace with z_goal-norm / residue-progression gate). Required to unblock the infant_substrate GAP-14 and the curriculum-vs-flat ARC-046 comparison.
5. MECH-341 entropy_bonus_scale retune + stratified trigger condition revision – substrate_queue entry added by governance today as a follow-on after the V3-EXQ-611 substrate-readiness landing. Parametric calibration of Q-054 entropy_lambda + stratified-temperature trigger threshold.
6. V3-EXQ-543k disposition recording (carry-forward) – still outstanding from 2026-05-21 drained-without-manifest BLOCK flagged by the governance verification gate; neither today’s failure-autopsy nor governance closes it.
7. Central runner_status freshness watch (resolved this cycle but worth monitoring) – the prior ~76h stall recovered to ~7h after governance staging. The Phase-2 coordinator -> central-index merge is catching up but the underlying intermittent wedge mode is what Phase-3 cutover is designed to mitigate.
8. Fleet contraction triage (carry-forward) – only DLAPTOP-4.local wrote within the last 24h. ree-cloud-3 is now ~42h stale (was ~18h yesterday); the other six hosts remain stale since 2026-05-21 or earlier. Decide whether to restart cloud workers or accept a Mac-only fleet through the MECH-341 + ARC-046 substrate-enrichment phase.
9. Coordinator Phase-3 go/no-go (carry-forward) – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon) and all writer review followups landed today (MED-A/B + LOW-A/C); the cluster is ready for the PHASE3_GIT_WRITER_READY=True flag flip pending the (still-open) operational question about hub power-loss durability.

---

Status Snapshot (2026-05-27T01:10Z – nightly docs sync; central runner_status static for ~76h post 2026-05-23T21:06Z (Phase-2 coordinator -> central-index merge longest stall of the cycle; only Mac + ree-cloud-3 still in active production); queue 2 -> 0 (entire 6-entry fix-and-retest cohort drained + V3-EXQ-590a/591 ISEF cohort pruned); pending_review.md regenerated 2026-05-25T09:06Z reads 0 but is stale (~5 fresh manifests landed since incl V3-EXQ-543l mixed + V3-EXQ-608 P2 PASS R2a + V3-EXQ-603/603a/603b + V3-EXQ-609); active governance cycle governance-20260526T230808Z opened 2026-05-26T23:08Z; no new SD/MECH/ARC/Q landings; ree-v3 substrate edits are MECH-111 dead-branch deletion + EXQ-571 per-channel instrumentation only; V3-EXQ-543k disposition recording outstanding; six 2026-05-26 IGW/diagnose-errors closures recorded)

• SDs / MECHs / claim landings since the 2026-05-25T01:10Z snapshot: none. Fourth consecutive nightly read with no substrate-side claim landing; the substrate state remains the 2026-05-21 wave + earlier (MECH-282 / MECH-286 / MECH-340 + Q-053 wiring + ARC-062 GAP-B mode-separation floor + coordinator Phase-3 cutover substrate). claims.yaml unchanged at 650 claims.
• ree-v3 substrate edits this window (instrumentation + dead-branch cleanup only):
  • MECH-111 dead-branch deletion 2026-05-25 (ree-v3 099743e) – ree_core/predictors/e3_selector.py:606-613 uniform-scalar-shift broadcast novelty branch dropped (argmin-invariant by construction; driver-verified bit-identical pre/post behaviour); pure-cleanup follow-on to the 2026-05-25 MECH-314a propagation root-cause investigation. claims.yaml MECH-111 untouched (no governance act, pure code-side cleanup).
  • EXQ-571 decomp per-channel std_across_K + bias_range_mean instrumentation 2026-05-25 (ree-v3 5c84a4d) – additive instrumentation keys on ree_core/agent.py surfacing per-candidate spread that the EXQ-571 mean-collapse metric obscured. Read by V3-EXQ-609 diagnostic. The same 2026-05-25 investigation identified the upstream E2 world-forward per-candidate z_world collapse (cand_world_pairwise_dist=0.0000 across K=32 candidates with 2-3 unique argmax classes) as the structural blocker on MECH-314a / MECH-320 / MECH-295 / SD-033a / SD-033b bias-channel diversity, captured in REE_assembly/evidence/planning/v3_exq_571_root_cause_2026-05-25.md.
  • REE_assembly-side governance dispositions landed against the V3-EXQ-571 finding (no claim landings): ARC-065 + MECH-314 failure_record[0].metric corrected in substrate_queue.json (683c252158); Q-044 blocked_by: [e2_world_forward_first_action_preservation] + dated blocker appendix added to notes (2eb5252f3f); MECH-314a Phase-2 novelty-source design doc + design_question entry in substrate_queue (5ec31e39c8 + 039e195637).
• 2026-05-26 IGW / diagnose-errors / governance closures (no code; disposition + planning only):
  • V3-EXQ-543l ARC-062 GAP-B falsifier successor PASS 2026-05-26T02:30:59Z manifest evidence_direction=mixed, claim_ids=[ARC-062, MECH-309, INV-074, MECH-334] – the first contributory-direction manifest in the ARC-062 cluster since the 2026-05-18 governance cycle set the prior 543f/g/h/i cohort to superseded. Governance disposition deferred to the in-flight cycle (governance-20260526T230808Z).
  • V3-EXQ-608 P2 PASS 2026-05-26T02:58:32Z manifest interpretation_label=R2a_e3_collapse_confirmed_large_gap across all 3 seeds (frac_pre_ge2=1.0, frac_e3_collapse_above_eps=0.858, mean_top2_class_gap=0.378) – triggered the MECH-341 Phase-3 substrate-design phase per the behavioral_diversity_isolation_plan.md decision rules. Closure-card frontmatter for the plan was retrofitted later the same day (REE_assembly cf3012570a, 8 nodes: GAP-A partial + GAP-B/C/D in_progress + GAP-E..H deferred).
  • Six IGW / diagnose-errors NO-FIX closes logged 2026-05-26: IGW-25 (V3-EXQ-250a SIGTERM-on-ree-cloud-1 with successors already reviewed), IGW-26 (V3-EXQ-538 SIGTERM with explicit in-script HOLD + 4 prereq boxes), IGW-21 (V3-EXQ-455a planning-stub subsumed by V3-EXQ-608 P2 PASS), IGW-30 + IGW-38 (ARC-062 retest stable-hash duplicates of in-flight V3-EXQ-543l), IGW-37 (cleared as stale ~29h, predated the IGW spawn fix). Stable-hash ledger entries recorded so subsequent hourly ticks stop re-spawning.
  • IGW hourly spawn fix landed 2026-05-26 (REE_Working 252494a, scripts/igw_routine_tick.py) – replaced screen-based spawn with claude -p --session-id <uuid> --dangerously-skip-permissions under setsid; orphan-branch cleanup helper + git worktree add timeout 60s -> 180s + TimeoutExpired recovery path. Zombie bash processes from IGW-37 / IGW-38 cleaned up. End-to-end verified at 17:14Z --force tick spawning IGW-20260526-030.
  • Workset Recent Activity panel added to /workset 2026-05-26 (REE_assembly 4fa044d8a9 serve.py + workset.html v2026-05-26.3) – new GET /api/workset/recent_activity?hours=N endpoint reading from igw_assignments.json released events; collapsible panel below the summary bar with hours dropdown + “recent N” chip + click-to-scroll behaviour. Serve.py restart needed for new endpoint.
  • MECH-307 / EXQ-571 dispositions (mech295 bridge config session): MECH-295 liking-bridge keeps the same defaults; Q-044 + ARC-065 + MECH-314 substrate_queue corrections + the MECH-314a Phase-2 design doc above are the five-disposition response to the EXQ-571 finding. None of these are claim landings; the design-question routing remains pending the user picking an option.
• Runner activity since the 2026-05-25T01:10Z snapshot: central evidence/experiments/runner_status.json reports 753 -> 753 cumulative completions (+0); all-time breakdown unchanged at 176 PASS / 303 FAIL / 81 ERROR / 193 UNKNOWN; last_updated 2026-05-23T21:06:24Z unchanged – the central file has now been static for ~76h (~3.2 days), the longest stall of the cycle and a worse state than yesterday’s ~52h. Cross-machine aggregate from per-host files: DLAPTOP-4.local 592 + ree-cloud-1 243 + ree-cloud-2 184 + ree-cloud-3 142 + ree-cloud-4 141 + ree-worker-3 133 + EWIN-PC 77 + Daniel-PC 28 = 1540 cumulative. Only DLAPTOP-4.local (heartbeat 2026-05-27T01:11Z, ~0h fresh) and ree-cloud-3 (heartbeat 2026-05-26T07:24Z, ~18h fresh) wrote within the last 24h; the other six hosts (incl ree-cloud-1/2/4 stale since 2026-05-21) appear effectively offline. Multi-machine fleet has contracted hard.
• Queue depth: 2 -> 0 items (queue file items: [] empty at this read). The full 6-entry fix-and-retest cohort queued 2026-05-25 (V3-EXQ-603a / 483d / 543l / 608 / 588b / 598b) has drained; V3-EXQ-543l + V3-EXQ-608 wrote manifests; the others’ fates will be resolved by the in-flight governance cycle. V3-EXQ-590a + V3-EXQ-591 (the EXQ-ISEF-004/005 cohort that was the load-bearing bottleneck for three snapshots) also no longer appear in the queue – the explicit-rescue checkpoint-resumable V3-EXQ-590a single-host pin appears to have been pruned without a fresh manifest, which warrants a separate diagnostic pass (V3-EXQ-590a was supposed to be checkpoint/resume-protected so an unexplained queue absence is concerning).
• Pending review: stale at 0 items (pending_review.md regenerated 2026-05-25T09:06:32Z; last review 2026-05-25T08:56:04Z). ~5 V3 manifests have landed in the intervening ~40h: V3-EXQ-543l (mixed against [ARC-062, MECH-309, INV-074, MECH-334]); V3-EXQ-608 (PASS R2a); V3-EXQ-603 / V3-EXQ-603a / V3-EXQ-603b (Q-045 / MECH-313 / MECH-260 retest cohort); V3-EXQ-609 (claim_ids=[] diagnostic). The active governance-20260526T230808Z cycle (started 2026-05-26T23:08Z, still listed status: active in TASK_CLAIMS.json) will walk these.
• Bottleneck note (significant reframe): the upstream E2 world-forward per-candidate z_world collapse identified by the 2026-05-25 V3-EXQ-571 root-cause investigation is now the structural root cause of the score_bias-chain flatness that has held ARC-065 (MECH-314a/b/c) + ARC-066 (MECH-320) + ARC-062 GAP-B (gated_policy heads) + MECH-295 / SD-033a/b at non_contributory across the diversity cluster. The finding doc enumerates four forward paths (Option A rolling z_world visitation buffer as recommended Phase-1 caveat; Option B first-action one-hot bypass per existing GAP-B fix; Option C candidate-pool relative rank; Option D hybrid harm + visitation; Option F action-object-identity at proposer stage surfaced during evaluation). ARC-062 / MECH-309 now has its first contributory-direction manifest in the cluster: V3-EXQ-543l PASS 2026-05-26T02:30Z evidence_direction=mixed (governance disposition deferred). V3-EXQ-608 P2 PASS confirmed R2.a “e3_collapse_confirmed_large_gap” on all 3 seeds and triggered the MECH-341 Phase-3 substrate-design phase. V3-EXQ-543k disposition recording (the 2026-05-21 drained-without-manifest BLOCK flagged by the governance verification gate) remains outstanding – 543l does not formally close 543k’s audit record. The night’s wrinkles: central runner_status freshness now at ~76h (worsening); fleet has contracted to Mac + cloud-3 active; the 590a/591 ISEF-004/005 cohort vanished from the queue without a fresh manifest path.

Immediate Work Queue (This Cycle, 2026-05-27)

1. governance-20260526T230808Z cycle outcome watch – the active governance cycle holds claims.yaml + claims.json + pending_review.md + review_tracker.json + substrate_queue.json + promotion_demotion_recommendations.md + experiment_proposals.v1.json. Its walk will resolve the ~5 new manifests landed since the 2026-05-25T09:06Z pending_review regen (V3-EXQ-543l mixed; V3-EXQ-608 PASS R2a; V3-EXQ-603/603a/603b cohort; V3-EXQ-609 diagnostic). Re-run generate_pending_review.py after the walk to confirm clear.
2. MECH-341 Phase-3 substrate-design routing decision – V3-EXQ-608 P2 PASS R2.a e3_collapse_confirmed_large_gap (2/3 majority across all 3 seeds, mean top-2 class gap 0.378) triggered the MECH-341 substrate-design phase per behavioral_diversity_isolation_plan.md. Use /implement-substrate to pick a design option (1/2/3) using the 608 large-gap evidence. Subsumes the SD-032a-isolation question that V3-EXQ-455a planning-stub was designed to test.
3. V3-EXQ-590a + V3-EXQ-591 disappearance investigation – both queue entries that were load-bearing for three snapshots are no longer in experiment_queue.json. V3-EXQ-590a was explicitly checkpoint/resume-protected and pinned ree-cloud-3 priority 100; its absence without a fresh manifest needs a coordinator / per-machine runner_status diagnostic pass. Re-queue if no manifest exists.
4. V3-EXQ-543k disposition recording – the 2026-05-21 drained-without-manifest BLOCK flagged by the governance verification gate is still open. V3-EXQ-543l is the active successor on the GAP-B mode-separation-floor axis; the audit gap is the 543k force_rerun whose remote fate was opaque. Record disposition in evidence/verification/ so the gate clears the historical incident.
5. Central runner_status freshness investigation (longest stall yet) – the central evidence/experiments/runner_status.json has been static for ~76h since 2026-05-23T21:06Z, the longest stall of the cycle. The Phase-2 coordinator -> central-index merge has been intermittently wedged across the last three snapshots and is now in its worst state. Per-machine runner_status/<hostname>.json files carry live writes for Mac + ree-cloud-3 but the central aggregator is not catching up. Operational hygiene priority before the next governance walk.
6. Fleet contraction triage – only DLAPTOP-4.local (~0h fresh) and ree-cloud-3 (~18h fresh) wrote within the last 24h. ree-cloud-1/2/4 last wrote 2026-05-21 (~6 days), Daniel-PC 2026-04-10 (~7 weeks), EWIN-PC 2026-04-21 (~5 weeks), ree-worker-3 2026-05-17 (~10 days). Decide whether to restart the cloud workers or accept a Mac-plus-cloud-3-only fleet for the duration of the diversity-cluster substrate-design phase.
7. MECH-314a Phase-2 novelty-source design-question decision – substrate_queue entry MECH-314a-Phase-2 (status=design_question, ready=false, priority=2) holds four candidate options (A rolling z_world visitation buffer; B first-action one-hot bypass per GAP-B; C candidate-pool relative rank; D hybrid harm + visitation). Companion design doc landed 2026-05-26 (REE_assembly 5ec31e39c8) recommended Option A as the Phase-1-caveat re-spec of MECH-314a signal source. User decision required before the substrate option lands.
8. Coordinator Phase-3 go/no-go (carry-forward) – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon); Phase-2 coordinator claims remain authoritative until a deliberate cutover decision. The current ~76h central-index stall is exactly the failure mode Phase-3 is intended to mitigate; worth pairing the decision with item 5.

---

Status Snapshot (2026-05-25T01:10Z – nightly docs sync; central runner_status static again (no completions in central index since 2026-05-23T21:06Z, now ~52h stale – merge re-stalled); queue depth unchanged at 2 (V3-EXQ-591 claim record day-old without movement – worth a manual check); pending_review 6 -> 0 (all 4 FAIL + 2 ERROR governance-walked as non_contributory or fix-and-requeued); 6 fix-and-retest queue-experiment sessions launched today (603a / 483d / 543l / 608 / 588b / 598b); V3-EXQ-606b held; closure-map plan frontmatter sync; no new SD/MECH/ARC/Q landings)

• SDs / MECHs / claim landings since the 2026-05-24T01:10Z snapshot: none. Third consecutive nightly read with no substrate-side change; the substrate state remains the 2026-05-21 wave + earlier (MECH-282 / MECH-286 / MECH-340 + Q-053 wiring + ARC-062 GAP-B mode-separation floor + coordinator Phase-3 cutover substrate). claims.yaml unchanged at 650 claims.
• Governance / queue-staging activity since the 2026-05-24T01:10Z snapshot (substantial, no code):
  • All 6 pending items closed as non_contributory via the failure-autopsy + diagnose-errors sessions opened yesterday evening:
    • V3-EXQ-603 (Q-045 / MECH-313 / MECH-260) – interactive Step 8 gate 2026-05-24T19:15Z confirmed all three claims NC. MECH-260 “weakens” is a call-path measurement artefact (act_with_split_obs() bypasses select_action() where dacc.record_action() fires; FIFO permanently empty; ARM_2==ARM_0 to 6 d.p. across 2 independent runs). Manifest overrides + claims.yaml evidence_quality_note additions are open under the next /governance cycle.
    • V3-EXQ-483c (SD-037 / MECH-280 / MECH-281 dACC measurement gap) – use_dacc=True omitted from all 4 GAP-4 arm configs; agent.dacc is None; _dacc_bias_norm() returns 0.0; C2 cannot fire. Cluster: 5 GAP-4 tier-1 experiments share this structural gap. SD-037 “weakens” tag is not scientific signal.
    • V3-EXQ-597b (MECH-258 PE-vs-raw post-SP-CEM) – dacc_suppression_weight=4.0 dominated pre-clip bias (~10-12), diluting PE signal; post-clip constant at 2.0 eliminated behavioural contrast. MECH-258 intact (not weakened; claim not fairly tested).
    • V3-EXQ-606a (ARC-064 GAP-I MECH-318 multi-rule empirical gate) + V3-EXQ-598 (SD-033a bias-head trainable ablation) – both ERROR root-caused to a git-sync gap on ree-cloud-2 (last pull 2026-05-21T13:18Z, scripts pushed 2026-05-21T14:37Z). No code bugs.
  • Six fix-and-retest queue-experiment sessions launched (still active at this read, all 2026-05-24T10:05Z-10:53Z window):
    • V3-EXQ-603a (priority 7; supersedes V3-EXQ-603) – repaired Q-045/MECH-313/MECH-260 four-arm ablation switching from act_with_split_obs() to select_action() training loop + obs_harm_a + use_affective_harm_stream + FIFO_WARMUP_STEPS=75. validate_queue + dry-run OK.
    • V3-EXQ-483d – SD-037 broadcast GAP-4 tier-1 retest with PAG/override_signal C2 criterion and goal_norm_peak C3 metric.
    • V3-EXQ-543l (supersedes 543k) – ARC-062 GAP-B mode-separation falsifier successor with escalated MODE_SEPARATION_FLOOR 0.25 -> 0.5 and P1_W_DEVIATION_AUX_WEIGHT 0.1 -> 0.3.
    • V3-EXQ-608 – MECH-319 GAP-K simulation-mode rule-write-gate falsifier with explicit dream replay loop, lateral_pfc-only (no gated_policy), three arms: admit_writes=False / True / gate-OFF baseline.
    • V3-EXQ-588b – goal-seeding pipeline diagnostic (post-588 autopsy), claim_ids=[], supersedes V3-EXQ-588.
    • V3-EXQ-598b (supersedes 598a; commitment_closure:GAP-1) – copy of 598a with GATES_ON_EXQ fixed from V3-EXQ-543k to V3-EXQ-543l (543k FAILed 2026-05-21; 543l is the new active successor).
  • V3-EXQ-606b held off the queue per user 2026-05-24T16:06Z (HOLD disposition for IGW-20260521-004). Gate dependency V3-EXQ-543k FAILed 2026-05-22 (evidence_direction=mixed; ARC-062=weakens, MECH-309=supports); no contributory 543-successor exists in evidence/ or queue as of this read. 606b script remains on disk but NOT queued; runtime gate would hard-block any run on the failed 543k manifest.
  • Closure-map audit + plan-frontmatter sync 2026-05-24T16:20Z – serve.py CLOSURE_KNOWN_PLANS +2 (arc_062_rule_apprehension, infant_substrate – both had frontmatter but were invisible in /closure tab); plan frontmatter synced across today’s queuing sessions (arc_062 GAP-B owner 543k -> 543l, GAP-H owner 603 -> 603a, GAP-K owner TBD -> 608, GAP-C/D substrate notes 543k -> 543l + 598 -> 598b; commitment_closure GAP-1 owner 598 -> 598b gating on 543l; sleep_substrate GAP-2 resume_condition 543b/c -> 543l; goal_pipeline GAP-4 resume records 483c FAIL NC + 483d queued).
• Runner activity since the 2026-05-24T01:10Z snapshot: central evidence/experiments/runner_status.json reports 753 -> 753 cumulative completions (+0); all-time breakdown unchanged at 176 PASS / 303 FAIL / 81 ERROR / 193 UNKNOWN; last_updated 2026-05-23T21:06:24Z unchanged – the central file is now ~52h stale at this read, having stalled again after one brief catch-up read yesterday. The 6 fix-and-retest queue entries listed above were staged today but have not yet returned manifests.
• Queue depth: 2 -> 2 items (unchanged at the surface; the 6 newly queued fix-and-retest entries above are in the queue but the surface count likely lags depending on whether the runner has pruned the pre-existing entries). V3-EXQ-590a (EXQ-ISEF-004 novelty-bonus Goldilocks calibration, MECH-314, checkpoint-resumable, pinned ree-cloud-3, priority 100; partial 1/15 runs saved; unclaimed at this read) + V3-EXQ-591 (EXQ-ISEF-005 4-phase infant curriculum vs flat baselines, ARC-046, claim record DLAPTOP-4.local 2026-05-23T21:06:24Z unchanged for ~28h – the claim record has not moved since yesterday’s snapshot; worth a manual check whether the claim has gone stale on the local runner).
• Pending review: 6 -> 0 items (pending_review.md regenerated 2026-05-24T19:20:42Z; last review 2026-05-24T18:16:00Z). All 4 FAIL + 2 ERROR from yesterday’s snapshot are now governance-cleared; the autopsy + diagnose-errors fixes are queued for retest as above.
• Bottleneck note (unchanged framing with one new wrinkle): EXQ-ISEF-004/005 (V3-EXQ-590a + 591) remain the parallel load-bearing developmental warm-start gate for the ARC-065 diversity narrative and the deferred Q-043/044/045 + INV-049 retests; the 591 day-old unchanged claim record is the new wrinkle – if DLAPTOP-4.local is not actively running it, the experiment needs to be released and re-claimed. ARC-062 / MECH-309 stays substrate_ceiling-framed; the new V3-EXQ-543l force-escalated GAP-B falsifier successor is the active test (queue priority TBD; gated on the runner picking it up). The previous-snapshot V3-EXQ-543k drained-without-manifest BLOCK is now superseded by 543l queue-staging at the plan-frontmatter level (the governance verification gate will still flag 543k as a historical incident until its disposition is formally recorded in evidence/verification/). The night’s good news: 0 pending review for the first time since 2026-05-22T05:30Z; bad news: central runner_status has been static for ~52h and V3-EXQ-591 has not advanced in ~28h.

Immediate Work Queue (This Cycle, 2026-05-25)

1. V3-EXQ-591 claim freshness check – the DLAPTOP-4.local 2026-05-23T21:06Z claim record on V3-EXQ-591 has been unchanged for ~28h. Confirm via runner heartbeats / serve.py /machines whether the experiment is actively running or whether the claim has been silently abandoned; release + re-queue if stale.
2. V3-EXQ-590a outcome watch (EXQ-ISEF-004) – novelty-bonus Goldilocks calibration with checkpoint/resume; ree-cloud-3 only (partial 1/15 runs, ARM_nbw01). Priority 100. Do not run on other hosts. PASS gates MECH-314 + Q-043/044.
3. V3-EXQ-543l outcome watch (ARC-062 GAP-B successor) – force-escalated mode-separation-floor falsifier (floor 0.25 -> 0.5; aux 0.1 -> 0.3). A contributory PASS unblocks V3-EXQ-598b (commitment_closure:GAP-1) + V3-EXQ-606b (held); a FAIL escalates ARC-062 / MECH-309 toward ARC-063 / V4 distributed CandidateRule field. One of the most-watched experiments in flight.
4. V3-EXQ-603a / 483d / 588b / 608 outcome watches – the fix-and-retest cohort queued today. Each addresses a fairness-of-test gap (call-path / config / etc.), not a substrate-ceiling claim; PASS or FAIL both contribute to the parent claims’ evidence. V3-EXQ-608 in particular is the first behavioural test of MECH-319 simulation-mode write gating with the admit_writes=True falsifier control arm.
5. V3-EXQ-543k disposition recording – the 2026-05-21 drained-without-manifest BLOCK was previously flagged by the new governance verification gate. With V3-EXQ-543l now staged as the successor, record the 543k disposition in evidence/verification/ so the gate clears that historical incident.
6. Central runner_status freshness investigation (recurring) – the central file has been static for ~52h post the brief 2026-05-24T21:06Z catch-up. Per-machine runner_status/<hostname>.json files carry the live writes; the coordinator -> central-index merge that intermittently catches up may be wedged again. Worth a diagnostic pass before the morning governance walk.
7. Coordinator Phase-3 go/no-go – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon); Phase-2 coordinator claims remain authoritative until a deliberate cutover decision.

---

Status Snapshot (2026-05-24T01:10Z – nightly docs sync; central runner_status freshness restored to ~4h (was ~35h); +4 completions (+2 FAIL +2 ERROR); pending_review 10 -> 6 (yesterday’s 5 substrate PASSes now reviewed); queue depth unchanged at 2; governance verification gate + shadow stale-TTL fix landed 2026-05-23; failure-autopsy + diagnose-errors sessions opened on the 4 FAIL + 2 ERROR pending items; no new SD/MECH/ARC/Q landings)

• SDs / MECHs / claim landings since the 2026-05-23T01:10Z snapshot: none. The substrate state is unchanged from yesterday; the 2026-05-21 wave (MECH-282 / MECH-286 / MECH-340 + Q-053 wiring + ARC-062 GAP-B mode-separation floor + coordinator Phase-3 cutover substrate) is now substrate-side stable and its 5 substrate-readiness PASSes are governance-cleared. claims.yaml unchanged (650 claims).
• Infrastructure / governance tooling since the 2026-05-23T01:10Z snapshot:
  • Governance verification gate landed 2026-05-23T20:38Z (REE_assembly master a686bfbc66) – new scripts/verify_governance_cycle.py + scripts/generate_governance_handoff.py + evidence/verification/ + evidence/handoffs/ + docs/governance_verification_gate.md. Additive safety layer with 7 checks (governance-cycle freshness, runner_status freshness, queue claim ages, etc.); exits 0/1/2 and writes JSON report + Markdown handoff. Current gate result: FAIL with 1 BLOCK (V3-EXQ-543k drained without manifest) and 2 WARN (the 54h central-runner_status staleness that has since resolved; ree-cloud-1 heartbeat divergence).
  • Shadow coordinator stale-machine TTL fix 2026-05-23T21:15Z (REE_assembly master 418c79ca4f) – 6h TTL filter on read_machines() and read_shadow_status() in serve.py so stale machines are excluded from both the git-backed machines dashboard and the coordinator shadow coordination view. Configurable via MACHINE_STALE_EXCLUDE_HOURS. Restart serve.py to activate.
  • Active failure-autopsy + diagnose-errors sessions opened 2026-05-23T22:14-22:36Z against the 4 FAIL + 2 ERROR pending items: failure-autopsy of V3-EXQ-597b (MECH-258 PE-vs-raw post-SP-CEM measurement gap), V3-EXQ-483c (SD-037 / MECH-280 / MECH-281 dACC measurement gap), V3-EXQ-603 (Q-045 / MECH-313 / MECH-260 measurement gap); /diagnose-errors re-queue of V3-EXQ-606a -> V3-EXQ-606b (ARC-064 GAP-I MECH-318 multi-rule empirical gate) and V3-EXQ-598 -> V3-EXQ-598a (SD-033a bias-head trainable ablation).
• Runner activity since the 2026-05-23T01:10Z snapshot: central evidence/experiments/runner_status.json reports 749 -> 753 cumulative completions (+4); all-time breakdown 176 -> 176 PASS (+0) / 301 -> 303 FAIL (+2) / 79 -> 81 ERROR (+2) / 193 UNKNOWN (unchanged); last_updated 2026-05-21T14:26:59Z -> 2026-05-23T21:06:24Z – the central file freshness has recovered to ~4h at this read (was ~35h yesterday), so the Phase-2 coordinator -> central-index merge is catching up rather than wedged. The +4 completions are the new pending-review entries (V3-EXQ-603 second run FAIL + V3-EXQ-606a / 598 ERROR + V3-EXQ-597b FAIL).
• Queue depth: 2 -> 2 items (unchanged). V3-EXQ-590a (EXQ-ISEF-004 novelty-bonus Goldilocks calibration, MECH-314, checkpoint-resumable, pinned ree-cloud-3, priority 100; partial 1/15 runs saved; unclaimed at this read) + V3-EXQ-591 (EXQ-ISEF-005 4-phase infant curriculum vs flat baselines, ARC-046, claimed DLAPTOP-4.local 2026-05-23T21:06Z).
• Pending review: 10 -> 6 items (pending_review.md regenerated 2026-05-23T22:03:10Z; last review 2026-05-23T21:57:44Z). The 5 substrate PASSes from yesterday’s snapshot (V3-EXQ-601 x2 MECH-269b-followup-A staleness gate, V3-EXQ-599a MECH-286 override-gated sleep onset, V3-EXQ-600a MECH-282 LPB interoceptive routing, V3-EXQ-607 MECH-340 persistence/efficacy gate + Q-053) have been moved into reviewed_run_ids – the entire 2026-05-21 substrate-validation diagnostic battery is now fully governance-cleared. Remaining: 4 FAIL (V3-EXQ-483c SD-037 broadcast-override GAP-4 tier-1; V3-EXQ-597b MECH-258 PE-vs-raw post-SP-CEM; V3-EXQ-603 x2 Q-045 MECH-313/MECH-260 collapse falsifier). 2 runner-only ERROR (V3-EXQ-606a ARC-064 GAP-I MECH-318 multi-rule empirical gate; V3-EXQ-598 SD-033a bias-head trainable ablation). The failure-autopsy + diagnose-errors sessions opened tonight cover all 6.
• Bottleneck note (unchanged framing): EXQ-ISEF-004/005 (V3-EXQ-590a + 591) remain the parallel load-bearing developmental warm-start gate for the ARC-065 diversity narrative and the deferred Q-043/044/045 + INV-049 retests. ARC-062 / MECH-309 stays substrate_ceiling-framed; the V3-EXQ-543k drained-without-manifest BLOCK is now formally flagged by the new governance verification gate (no longer just an opaque queue absence). The night’s good news is on the operational hygiene side: central runner_status freshness restored, governance verification gate live as a hard tripwire on future drained-without-manifest incidents, shadow-coordination dashboard filtered for stale machines.

Immediate Work Queue (This Cycle, 2026-05-24)

1. Failure-autopsy outcomes (V3-EXQ-597b, V3-EXQ-483c, V3-EXQ-603) – 3 active failure-autopsy sessions opened 2026-05-23T22Z. Each is diagnosing whether the FAIL is a measurement gap (re-queue corrected) vs a true substrate-ceiling result (governance reclassify as non_contributory). Review the autopsy artifacts (evidence/planning/failure_autopsy_V3-EXQ-*_2026-05-23.{md,json}) on completion.
2. diagnose-errors re-queue (V3-EXQ-606b, V3-EXQ-598a) – ERROR -> fix re-queue active. 606b addresses the ARC-064 GAP-I MECH-318 multi-rule empirical gate; 598a addresses the SD-033a bias-head trainable ablation. Both depend on V3-EXQ-543k contributory PASS; watch for queue-policy violations.
3. V3-EXQ-543k re-queue decision – the drained-without-manifest BLOCK is now flagged by the governance verification gate as the longest-standing unresolved item. Either re-queue with force_rerun (third attempt) or governance-record as a substrate-side dead end and route ARC-062 / MECH-309 toward ARC-063 / V4 distributed CandidateRule field.
4. V3-EXQ-590a outcome watch (EXQ-ISEF-004) – novelty-bonus Goldilocks calibration with checkpoint/resume; ree-cloud-3 only (partial 1/15 runs, ARM_nbw01). Priority 100. Do not run on other hosts.
5. V3-EXQ-591 outcome watch (EXQ-ISEF-005) – 4-phase infant curriculum vs flat baselines (GAP-14), claimed DLAPTOP-4.local 2026-05-23T21:06Z; PASS gates ARC-046 and DEV-NEED-008.
6. Governance walk the 6 pending after autopsy / diagnose-errors lands – re-run generate_pending_review.py to confirm clear; then run the new scripts/verify_governance_cycle.py to record a fresh handoff with the 543k BLOCK resolution.
7. Coordinator Phase-3 go/no-go – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon); Phase-2 coordinator claims remain authoritative until a deliberate cutover decision.

---

Status Snapshot (2026-05-23T01:10Z – nightly docs sync; 2026-05-21 substrate-validation diagnostics PASS in turn (MECH-269b-followup-A x2 / MECH-286 / MECH-282 / MECH-340); central runner_status static at 749 (~35h stale post Phase-2 cutover); queue depth 3 -> 2 (543k drained); pending_review 0 -> 10 with governance cycle still in flight; no new substrate landings)

• SDs / MECHs / claim landings since the 2026-05-22T01:10Z snapshot: none. The substrate state is unchanged from yesterday; the 2026-05-21 wave (MECH-282 / MECH-286 / MECH-340 + Q-053 wiring + ARC-062 GAP-B mode-separation floor + coordinator Phase-3 cutover substrate) is now substrate-side stable and is being read by tonight’s pending-review queue, not extended.
• Runner activity since the 2026-05-22T01:10Z snapshot: central evidence/experiments/runner_status.json reports 749 cumulative completions UNCHANGED (176 PASS / 301 FAIL / 79 ERROR / 193 UNKNOWN); last_updated 2026-05-21T14:26:59Z – the central file is now ~35h stale (Phase-2 coordinator cutover; per-machine runner_status/<hostname>.json files carry the live writes). All the completions in the intervening 24h appear in pending_review.md but have not been merged into the central index.
• Queue depth: 3 -> 2 items. V3-EXQ-590a (EXQ-ISEF-004 novelty-bonus Goldilocks calibration, MECH-314, checkpoint-resumable, pinned ree-cloud-3, priority 100; partial 1/15 runs saved) + V3-EXQ-591 (EXQ-ISEF-005 4-phase infant curriculum vs flat baselines, ARC-046, claimed DLAPTOP-4.local). V3-EXQ-543k drained this window – the 2026-05-21T14:13Z force_rerun re-queue is no longer in the queue, but no fresh 543k manifest appears in pending_review.md either; its remote fate is opaque from the central indices pending the next governance walk.
• Pending review: 0 -> 10 items (pending_review.md regenerated 2026-05-22T05:31:18Z; last review 2026-05-22T05:30:55Z). 5 PASS (V3-EXQ-601 x2 MECH-269b-followup-A staleness gate at the default 0.4/0.5 thresholds; V3-EXQ-599a MECH-286 override-gated sleep onset; V3-EXQ-600a MECH-282 LPB interoceptive routing + SD-037 split recruitment; V3-EXQ-607 MECH-340 persistence/efficacy gate + Q-053 agent-side appraisal wiring – the entire 2026-05-21 substrate-validation diagnostic battery PASSed in turn). 3 FAIL (V3-EXQ-597b MECH-258 PE-vs-raw post-SP-CEM C2-telemetry revalidation; V3-EXQ-603 x2 Q-045 MECH-313/MECH-260 collapse falsifier). 2 runner-only ERROR (V3-EXQ-606a ARC-064 GAP-I MECH-318 multi-rule empirical gate; V3-EXQ-598 SD-033a bias-head trainable ablation). The active governance cycle governance-20260522T032251Z (started 2026-05-22T03:22:51Z, still listed as active in TASK_CLAIMS.json) has not yet walked these.
• Bottleneck note (unchanged framing): EXQ-ISEF-004/005 (V3-EXQ-590a + 591) remain the parallel load-bearing developmental warm-start gate for the ARC-065 diversity narrative and the deferred Q-043/044/045 + INV-049 retests. ARC-062 / MECH-309 stays substrate_ceiling-framed; the GAP-B mode-separation-floor falsifier V3-EXQ-543k force_rerun outcome is opaque from the central indices and waits on the next governance walk. The night’s good news is on the validation side: every substrate-readiness diagnostic queued out of the 2026-05-21 landings (MECH-269b-followup-A, MECH-286, MECH-282, MECH-340 + Q-053) cleared its acceptance criteria.

Immediate Work Queue (This Cycle, 2026-05-23)

1. Governance walk the 10 pending – the governance-20260522T032251Z cycle was opened ~22h ago and is still listed active; the 5 substrate-readiness PASSes (601 x2 / 599a / 600a / 607) can be moved into reviewed_run_ids after the customary one-line interpretation per claim; 597b / 603 x2 FAIL want their failure-autopsy / NC dispositions; 606a / 598 ERROR want the runner-only acknowledgement path. Re-run generate_pending_review.py after the walk to confirm clear.
2. V3-EXQ-543k force_rerun outcome resolution – the re-queued entry drained without a central manifest. Confirm via per-machine runner_status files / coordinator panel whether 543k ran on a remote and the manifest is awaiting sync, or whether the queue entry was pruned without a run. Re-queue if needed.
3. V3-EXQ-590a outcome watch (EXQ-ISEF-004) – novelty-bonus Goldilocks calibration with checkpoint/resume; ree-cloud-3 only (partial 1/15 runs, ARM_nbw01). Priority 100. Do not run on other hosts.
4. V3-EXQ-591 outcome watch (EXQ-ISEF-005) – 4-phase infant curriculum vs flat baselines (GAP-14), claimed DLAPTOP-4.local; PASS gates ARC-046 and DEV-NEED-008.
5. Central runner_status backfill – the central file has been static for ~35h post Phase-2 coordinator cutover. Worth investigating whether the coordinator -> central-index merge is intentionally lagged or whether it has wedged; downstream indexer / governance pipeline derive from it.
6. Coordinator Phase-3 go/no-go – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon); Phase-2 coordinator claims remain authoritative until a deliberate cutover decision.

---

Status Snapshot (2026-05-22T01:10Z – nightly docs sync; MECH-282 / MECH-286 / MECH-340 (+Q-053 wiring) substrate landings; ARC-062 GAP-B mode-separation floor; coordinator Phase-3 cutover substrate designed; V3-EXQ-543k GAP-B falsifier re-queued; pending_review 0 with a governance cycle in flight)

• SDs / MECHs / claim landings since the 2026-05-21T12:39Z snapshot:
  • MECH-282 LPB interoceptive routing 2026-05-21 (ree-v3) – LPBInteroceptiveRouter splits harm into z_harm (external; resource slice zeroed before HarmEncoder) and a non-trainable z_harm_intero broadcast (drive_level + harm_obs_a resource EMA); SD-037 coupling routes intero -> override, external -> PAG freeze proxy when both flags on. use_lpb_interoceptive_routing default False. V3-EXQ-600 3-arm diagnostic queued.
  • MECH-286 override-gated sleep onset 2026-05-21 (ree-v3) – wake-stability axis of SD-037: the wake->offline transition is gated by a joint permit (override_signal below threshold AND max region-staleness above recruit AND z_harm_a.norm() below tonic). use_mech286_sleep_onset_gate default False preserves deterministic K-episode firing. V3-EXQ-599 3-arm diagnostic queued.
  • MECH-340 persistence/efficacy gate + Q-053 agent wiring 2026-05-21 (ree-v3) – ARC-079 / Q-053 front-runner: a GhostGoalBank entry’s persistence as a MECH-293 re-probe target is gated; disengagement is the default when license = control_efficacy * (1 - goal_unattainability) < persistence_floor (SD-039 trace preserved). Q-053 wiring maps prior hippocampal completion + E3 commitment -> control_efficacy and one-shot 1 - goal_proximity -> goal_unattainability. V3-EXQ-607 diagnostic queued; contracts 8/8 + dry-run PASS.
  • ARC-062 GAP-B mode-separation floor 2026-05-20 (ree-v3) – composed gated bias becomes w*h0 + (1-w)*h1 + floor*(h0-h1), injecting a non-cancelable mode contrast at discriminator w~0.5 (the V3-EXQ-543i autopsy gap where delta_hat cancels in base + (2w-1)*delta_hat). GatedPolicyConfig.mode_separation_floor default 0.0 – bit-identical OFF. Validation V3-EXQ-543k (supersedes 543i; 12-arm + floor/aux on gated arms, K=3 basin-stability gate) re-queued with force_rerun after a ree-cloud-4 FAIL with no central manifest (failure autopsy: a 543i manifest mis-filed under the 543k slot).
  • Coordinator Phase-3 cutover substrate designed 2026-05-21 (ree-v3) – PHASE3_CUTOVER.md + phase3_preflight / phase3_verify + sync_daemon scaffold + phase3_cutover.sh + tests; serve API; smoke PASS. Phase-2 (coordinator claims) remains live; Phase 3 is designed-not-cut-over.
• Runner activity since the 2026-05-21T12:39Z snapshot: central evidence/experiments/runner_status.json reports 736 -> 749 cumulative completions (+13); all-time breakdown 172 -> 176 PASS (+4) / 294 -> 301 FAIL (+7) / 77 -> 79 ERROR (+2) / 193 UNKNOWN (unchanged); last_updated 2026-05-21T14:26:59Z (central file lags per-machine writes). V3-EXQ-543k completed since the snapshot and is re-queued with force_rerun for a clean retest.
• Queue depth: 2 -> 3 items. V3-EXQ-590a (EXQ-ISEF-004 novelty-bonus Goldilocks calibration, MECH-314, checkpoint-resumable, pinned ree-cloud-3, priority 100; partial 1/15 runs saved) + V3-EXQ-543k (ARC-062 GAP-B mode-separation-floor falsifier, force_rerun, unclaimed) + V3-EXQ-591 (EXQ-ISEF-005 4-phase infant curriculum vs flat baselines, ARC-046, claimed DLAPTOP-4.local). Off-queue this window: Q-043/044/045 EXQs 603/604/605 (604/605 FAIL); V3-EXQ-597b / 598 / 606a queued and partially run.
• Pending review: 8 -> 0 items (pending_review.md regenerated 2026-05-21T13:02:46Z; last review 2026-05-21T12:59:23Z) – the 2026-05-21 governance walks cleared 543j/543i/595/597/598/604/605 as non_contributory and acked 599/600 ERROR. Caveat: +13 runner completions (including V3-EXQ-543k) have landed since that regeneration and a fresh governance cycle (governance-20260522T032251Z) is active at this read – the 0 is stale, the in-flight cycle will resolve the new manifests.
• Bottleneck note (unchanged framing): EXQ-ISEF-004/005 completion (V3-EXQ-590a + 591) remains the parallel load-bearing developmental warm-start gate for the ARC-065 diversity narrative and the deferred Q-043/044/045 + INV-049 retests. ARC-062 / MECH-309 stays substrate_ceiling-framed; the GAP-B mode-separation-floor falsifier V3-EXQ-543k is the immediate decisive test (re-queued, in flight). Coordinator Phase-3 cutover is designed but not executed – Phase-2 coordinator claims remain authoritative.

Immediate Work Queue (This Cycle, 2026-05-22)

1. V3-EXQ-543k outcome watch (ARC-062 GAP-B, load-bearing) – mode-separation-floor falsifier, supersedes 543i, re-queued with force_rerun. A contributory PASS unblocks V3-EXQ-598 (SD-033a bias-head ablation) and V3-EXQ-606a (ARC-064 GAP-I MECH-318 multi-rule gate); a FAIL holds ARC-062 / MECH-309 at substrate_ceiling and routes toward ARC-063 / V4.
2. V3-EXQ-590a outcome watch (EXQ-ISEF-004) – novelty-bonus Goldilocks calibration with checkpoint/resume; ree-cloud-3 only (partial 1/15 runs, ARM_nbw01). Priority 100. Do not run on other hosts.
3. V3-EXQ-591 outcome watch (EXQ-ISEF-005) – 4-phase infant curriculum vs flat baselines (GAP-14), claimed DLAPTOP-4.local; PASS gates ARC-046 and DEV-NEED-008.
4. Governance cycle in flight – governance-20260522T032251Z is walking the post-snapshot manifests (V3-EXQ-543k + the +13 completions since pending_review last regenerated); re-run generate_pending_review.py to confirm clear after it lands.
5. MECH-282 / MECH-286 / MECH-340 validation watch – V3-EXQ-599 / 600 / 607 substrate diagnostics queued; review on completion.
6. Coordinator Phase-3 go/no-go – Phase-3 cutover substrate designed (PHASE3_CUTOVER.md + preflight/verify + sync_daemon); Phase-2 coordinator claims remain authoritative until a deliberate cutover decision.

---

Status Snapshot (2026-05-21T12:39Z – roadmap Immediate Work Queue refresh; coordinator Phase-2 cutover + fleet pause; V3-EXQ-590a partial rescue; V3-EXQ-543j governance-closed non_contributory; goal_pipeline:GAP-3 done; pending_review 8)

• Coordination / fleet since the 2026-05-19T01:10Z snapshot:
  • Coordinator Phase-2 cutover 2026-05-21 – hub COORDINATOR_MODE=coordinator + SYNC_MODE=coordinator; Mac + ree-cloud-1..4 on coordinator claims; explorer panel Phase 2 live.
  • Fleet pause + runner checkpoint/suspend 2026-05-21 – suspend_run / resume_run + runner_checkpoint.py for checkpoint_resumable EXQs; fleet pause queued via scripts/fleet_pause_runners.py (runners need per-host restart to load new code; pause holds queue after current run ends).
  • V3-EXQ-590a partial rescue 2026-05-21 – only checkpoint on ree-cloud-3 (1/15 runs, ARM_nbw01); affinity pinned ree-cloud-3, priority 100; supersedes V3-EXQ-590 (stdout false PASS, no manifest). Duplicate 590/590a on other hosts stopped.
  • V3-EXQ-543j governance closure 2026-05-21T00:03Z – classified non_contributory (failure_autopsy_543i); removes the 2026-05-19 load-bearing cross-machine adjudication item from the active queue.
• Plan-of-record / substrate closures since 2026-05-19:
  • goal_pipeline:GAP-3 DONE 2026-05-20 – V3-EXQ-582a PASS (drive_floor=0.9); MECH-306 sustained_drive_trace registered.
  • EXQ-ISEF-001 PASS 2026-05-19 – V3-EXQ-587 harm-gradient curriculum; Rank-1 warm-start gate met at experiment level.
  • IGW substrate-ready wave 2026-05-21 – MECH-204, MECH-307, SD-037, INF-ENV-001/004, ARC-062 Phase 1 reconciled in substrate_queue; V3-EXQ-602/601/606 and Q-043/044/045 EXQs 603-605 queued and partially executed (604/605 FAIL manifests on disk, pending governance).
• Runner activity since the 2026-05-19T01:10Z snapshot: central evidence/experiments/runner_status.json reports 723 -> 736 cumulative completions (+13); all-time breakdown 166 -> 172 PASS (+6) / 287 -> 294 FAIL (+7) / 77 ERROR / 193 UNKNOWN (unchanged); last_updated 2026-05-21T08:35:34Z (central file lags per-machine writes).
• Queue depth: 6 -> 2 items. V3-EXQ-590a (ISEF-004 novelty Goldilocks checkpoint resume, MECH-314, pinned ree-cloud-3, priority 100, unclaimed at snapshot) + V3-EXQ-591 (ISEF-005 curriculum-vs-flat, ARC-046, pending, full rerun). Drained since 2026-05-19: 543j, 582a, 592, 481b, and the ISEF-002/003 slots; Q-043/044/045 EXQs 603-605 ran off-queue (manifests await review).
• Pending review: 1 -> 8 items (pending_review.md regenerated 2026-05-21T12:30:03Z; last review 2026-05-21T09:55:30Z). FAIL: V3-EXQ-597 (MECH-258, held for /failure-autopsy), V3-EXQ-595. Unclaimed FAIL manifests: 543i, 598, 604, 605. Runner-only ERROR: 599, 600.
• Bottleneck note (shift): EXQ-ISEF-004/005 completion (590a + 591) is now the parallel load-bearing developmental gate for ARC-065 warm-start narrative and deferred Q-043/044/045 + INV-049 retests. ARC-062 / MECH-309 remains substrate_ceiling-framed (543i/598/543k) but 543j is closed – no longer blocking on cross-machine byte-identical confirmation. Coordinator Phase-2 operational hygiene (fleet pause, single-host 590a, claim release) is the immediate coordination risk.

Immediate Work Queue (This Cycle, 2026-05-21)

1. V3-EXQ-590a outcome watch (ISEF-004, load-bearing) – novelty Goldilocks weight sweep with checkpoint/resume; ree-cloud-3 only (partial 1/15 runs saved, ARM_nbw01); supersedes V3-EXQ-590 false PASS (no manifest). Priority 100. Do not run on other hosts. Resume after fleet pause clears and runners restart with checkpoint/suspend code.
2. V3-EXQ-591 outcome watch (ISEF-005) – 4-phase infant curriculum vs flat baselines (GAP-14); pending unclaimed; full rerun (~120 min, no _partial checkpoint). PASS gates ARC-046 and DEV-NEED-008.
3. Coordinator Phase-2 + fleet pause – hub coordinator claims authoritative; fleet pause issued for rollout; per-host runner restart needed for suspend/resume_run + heartbeat time fields. Monitor coordinator panel; release orphaned claims; keep 590a single-host.
4. Governance walk (8 pending) – absorb V3-EXQ-543j as non_contributory (2026-05-21T00:03Z); review V3-EXQ-597 (/failure-autopsy), 595 FAIL, unclaimed 543i/598/604/605 FAIL manifests, runner-only 599/600 ERROR. Re-run generate_pending_review.py to confirm clear.
5. ARC-062 / MECH-309 cluster – V3-EXQ-543k still in_flight; 543i + 598 on disk pending review. Substrate_ceiling framing holds; no further 543j cross-machine watch.
6. goal_pipeline:GAP-2 – V3-EXQ-514j interpretation_branch grid landed on main; queue/run V3-EXQ-514k successor (reef + MECH-307 + drive_floor=0.9) when fleet unpauses.
7. goal-disengagement biology-before lit-pull – prerequisite for deferred ARC-078 C3 abandon mechanism + trigger-Q registration (carry forward).
8. Shadow-coordinator go/no-go – Phase-2 live; git fallback remains; watch /shadow/divergence under real coordinator load before further cutover.

---

Status Snapshot (2026-05-19T01:10Z – nightly docs sync; ARC-062 GatedPolicy differential-heads robustness fix landed; 2026-05-18 governance cycle superseded the 543f/g/h crystallization-falsifier cluster -> 543i, epistemic_category=substrate_ceiling; retrieval-cue reframe + ARC-078/MECH-339 registered (650 claims); GAP-L socially-scaffolded rule-population lit-pull discharged; Phase 0-1 shadow experiment-coordinator built; heartbeat-autostash governance regen recovered; pending_review 1)

• SDs / MECHs / claim landings since the 2026-05-18T01:10Z nightly:
  • ARC-062 GatedPolicy differential-heads robustness fix landed 2026-05-18 (ree-v3). When use_differential_heads=True the two scoring heads are SYNTHESIZED as a shared trunk plus a candidate-axis-norm-pinned differential (head_0 = base + delta_hat, head_1 = base - delta_hat, delta_hat = differential_bias_scale * delta / (||delta||_K + 1e-8)), making the cross-machine head_0==head_1 collapse attractor a structural non-equilibrium (delta==0 unreachable – scale-invariant normalization zeroes the magnitude gradient; at w=0.5 d(gated)/dw = 2*delta_hat != 0 by the norm pin). GatedPolicyConfig.use_differential_heads default False = bit-identical pre-fix path. crystallize() freezes (base, delta, discriminator). Validation V3-EXQ-543i (supersedes V3-EXQ-543g + V3-EXQ-543h) FAILed branch e (MECH-309 supports / ARC-062 weakens; all 4 diff-ON gated arms 3/3 inert despite delta_hat L2 pinned to differential_bias_scale=0.1) on a SINGLE machine (Mac).
  • 2026-05-18T16:30Z governance cycle – confirmed failure_autopsy_V3-EXQ-543h applied to the ARC-062 / MECH-309 crystallization-falsifier cluster: 7 manifests (543f x4 / 543g / 543h x2) set evidence_direction=superseded by V3-EXQ-543i, epistemic_category=substrate_ceiling, pending_retest_after_substrate; 543h INV-074 / MECH-334 branch corrected to non_contributory. claims.yaml notes/flags applied to ARC-062 / MECH-309 / INV-074 / MECH-334 (narrow_supports: zero reliable trained-policy evidence); MECH-332 / MECH-334 hold_pending_v3_substrate approved; claims.json rebuilt (645). V3-EXQ-543i FLAGGED for /failure-autopsy (diff_on_escape=true substrate fix works, diff_off non-reproduction) and deliberately left as the single pending item pending cross-machine confirmation.
  • retrieval-cue reframe + cue-system cluster registered 2026-05-18/19. The litpull-cue-system synthesis (9 biology entries into targeted_review_ghost_goal_search) drove an interpretive recast of ARC-060 / MECH-292 / MECH-293 / SD-039 as a content-addressed cued-retrieval system (status/phase/confidence unchanged – NO new claim/experiment/promotion in the reframe pass). Then ARC-078 (parent: unresolved-goal bank = content-addressed cue-addressed retrieval system satisfying C1-C3; ARC-060 specializes into it) + MECH-339 (C1 composite retrieval cue z_goal + SD-039 context channel, outshining gate) registered candidate / v3_pending. The C3 abandon mechanism + trigger Q DELIBERATELY NOT registered – gated behind a goal-disengagement biology-before lit-pull (active session; avoids SD-003-class philosophy-right/mechanism-wrong; mirror of the unregistered Zeigarnik gap, must be wired not parallel-registered). claims.yaml 648 -> 650 claims; validates (71 invariants).
  • GAP-L socially-scaffolded rule-population lit-pull discharged 2026-05-18 – 8 entries (Csibra & Gergely natural pedagogy, Wood/Bruner/Ross + Vygotsky ZPD, Tomasello shared intentionality, Tulving & Thomson encoding specificity, Godden & Baddeley + 2021 replication FAILURE, Nakazawa CA3 pattern completion, + Heyes 2016 / Spelke & Kinzler counterweights) in a new cross-cutting dir; lit_conf parallel signal on ARC-077 / MECH-337 / MECH-338 (NOT blended into exp_conf; all REMAIN candidate; the caregiver/teacher-agent substrate SECOND HARD GATE stays open).
• Infrastructure / coordination wave (2026-05-18 -> 2026-05-19):
  • Phase 0-1 shadow experiment-coordinator built (ree-v3/coordinator/, commit 67b9a73). Stdlib http.server + sqlite3 atomic-claim service (BEGIN IMMEDIATE conditional UPDATE = the mutex) + env-gated best-effort shadow client shim + additive shadow hooks in the runner. DEFAULT COORDINATION_MODE=git is a HARD no-op – the live runner is byte-identical to before; git stays authoritative; NO cutover. Local e2e 13/13 PASS (12-thread concurrent exactly-one-winner + idempotent /result replay + injected-divergence detection).
  • heartbeat-autostash governance-regen recovery – the 2026-05-18T21:27Z launchd git-sync repair (fetch hung ~55min then a safe-path reset --hard ate a dangling autostash because the dirty-check is blind to autostashed work) discarded a 639-file governance regen; the freshest 9 substantive governance-decision files (promotion_demotion 21:08Z incorporating ARC-063/077/MECH-337/338, conflicts, evidence_backlog, experiment_proposals(+index), architecture_gap_register(+md), decision_state, TODOs) restored from the dangling commit e154c853; the ~600 derivable per-run INDEX.md left for the next routine rebuild. Root cause (heartbeat autostash-rebase + un-timeout’d repair fetch + dirty-check blind spot) is mitigated, not yet fixed. Companion landings: rebase-free push_heartbeat / push_commands + cloud systemd repair timer (ree-v3 8be9b39); architecture-diagrams workflow fully green (claims.yaml dropped from the freshness trigger + mmdc –no-sandbox + render auto-commit rebase-retry).
  • All runners gracefully drained + cloud-2 force-stopped for a multi-machine coordination update; cloud-3’s unique untracked V3-EXQ-543g FAIL manifest preserved to origin before the drain. Cloud-3 push-to-origin livelock (hel1 latency vs 4 concurrent heartbeat-pushers) flagged for a user decision – duplicate compute, NOT data loss.
• Runner activity since the 2026-05-18T01:10Z nightly: central evidence/experiments/runner_status.json reports 720 -> 723 cumulative completions (+3); all-time breakdown 164 -> 166 PASS (+2) / 286 -> 287 FAIL (+1) / 77 ERROR / 193 UNKNOWN (unchanged); last_updated 2026-05-18T15:44:20Z – the central file is ~9.5h stale at this read (the multi-machine runners write per-machine runner_status/<hostname>.json).
• Queue depth: 11 -> 6 items. The ISEF cohort drained partially; remaining: V3-EXQ-590 (ISEF-004 novelty Goldilocks weight sweep, MECH-314, claimed), 591 (ISEF-005 curriculum-vs-flat baselines, ARC-046), 481b (MECH-090 V_s commit-release re-issue), 582a (goal_pipeline:GAP-3 Option 2 drive_floor sweep, supersedes 582), 592 (commitment_closure:GAP-11 committed-mode curriculum pilot, MECH-090), V3-EXQ-543j (ARC-062 differential-heads byte-identical cross-machine confirmation of the 543i 19:10Z run, priority 5, pinned ree-cloud-4).
• Pending review: 0 -> 1 item (pending_review.md regenerated 2026-05-18T16:32:44Z; last review 2026-05-18T16:30:27Z). The lone pending is v3_exq_543i_arc062_differential_heads_falsifier_20260518T063711Z_v3 (ARC-062 / INV-074 / MECH-309 / MECH-334), flagged for /failure-autopsy by the 2026-05-18T16:30Z governance cycle and deliberately held pending the V3-EXQ-543j cross-machine confirmation rather than reviewed prematurely.
• Bottleneck note (shift): the ARC-062 / MECH-309 rule-apprehension cross-machine bistability is now the load-bearing gate. The 543g->h->i chain exists because the same config landed gating-ACTIVE on host-A but INERT on cloud-3 AND cloud-4 – head collapse is the common cross-machine attractor and differentiation was a rare lucky-basin escape. V3-EXQ-543j (byte-identical to 543i in all executable code; the ONLY varied factor is the host machine; pinned ree-cloud-4, a participant in the divergent INERT basin) is decisive either way: branch e CONFIRMS-543i -> clears the single-machine caveat, ARC-062 demotion + ARC-063 / V4 strong-reading governance can proceed; branch a CONTRADICTS -> 543i was itself a single-machine basin artifact, do NOT demote ARC-062, escalate to a >=2-machine matrix. The ARC-065 behavioural-diversity developmental warm-start failure (EXQ-ISEF cohort V3-EXQ-587..591) remains the parallel dominant blocker.

Immediate Work Queue (This Cycle, 2026-05-19)

1. V3-EXQ-543j outcome watch (load-bearing) – ARC-062 differential-heads byte-identical cross-machine confirmation of the 543i 19:10Z run, pinned ree-cloud-4. Branch e CONFIRMS-543i -> ARC-062 demotion + ARC-063 / V4 governance proceed; branch a CONTRADICTS -> 543i was a single-machine basin artifact, escalate to a >=2-machine matrix + /failure-autopsy. The lone pending_review item (543i) is held until this lands.
2. goal-disengagement biology-before lit-pull – active session; new cross-cutting dir targeted_review_goal_disengagement (adaptive disengagement + trigger / failure-to-disengage rumination-OCD / over-disengagement apathy-avolition). Prerequisite for the deferred ARC-078 C3 abandon mechanism + trigger-Q registration.
3. EXQ-ISEF cohort outcome watch – V3-EXQ-590 (ISEF-004 novelty Goldilocks weight sweep, claimed), 591 (ISEF-005 curriculum-vs-flat baselines). The ARC-065 warm-start curriculum gate still blocks the deferred Q-043/044/045 + INV-049 retests.
4. goal_pipeline:GAP-3 – V3-EXQ-582a drive_floor insatiability sweep (Option 2 escalation from the V3-EXQ-582 substrate-ceiling FAIL); PASS closes GAP-3.
5. commitment_closure:GAP-11 – V3-EXQ-592 committed-mode curriculum pilot (EMERGENT / FORCED_RV / STARVED arms); PASS unblocks the OCD battery V3-EXQ-460b/463b/464b/466b/467b/468b.
6. Governance walk – next governance / morning-digest session should absorb the ARC-078 / MECH-339 + ARC-077 / MECH-337 / MECH-338 registrations into the recommendation queue, walk the V3-EXQ-543j result, and re-confirm pending_review after the cross-machine adjudication.
7. Shadow-coordinator go/no-go – watch /shadow/divergence over days of real load before any decision to advance past Phase 1; git stays authoritative until then.

---

Status Snapshot (2026-05-18T01:10Z – nightly docs sync; plasticity-crystallization cluster registered (INV-074/075 + MECH-333/334/335/336 + ARC-075/076 + Q-052; 645 claims); INV-074/MECH-333/334 Phase-3 crystallization substrate landed; ARC-062 GAP-B/C/D wiring; ARC-065 SP-CEM flipped to main-path default; SD-012 sustained-drive Option 1+2; commitment_closure GAP-3/GAP-11; infant_substrate EXQ-ISEF-001..005 cohort queued; pending_review 0)

• SDs / MECHs / claim landings since the 2026-05-17T01:11Z nightly:
  • Plasticity-crystallization cluster registered 2026-05-17 – INV-074 + MECH-333 + MECH-334 + ARC-075 + Q-052 (developmental critical-period crystallization: open-phase plasticity injection + closure write-protect); INV-075 (signal-structure temporal decoupling – a self-extinguishing AND load-bearing signal necessarily requires lock OR handoff; disjunction exhaustive; reframed as PARENT of INV-074 which becomes the competitive special case); ARC-076 + MECH-335 + MECH-336 (developmental window-then-lock calibration of the 3 BG commitment loops = personality; INV-075 LOCK-arm instance; two-layer L1 endogenous-timer set-points + L2 environmentally-calibrated commitment-policy) with an explicit implementation_prerequisites HARD ORDERING GATE (ARC-076/MECH-335/336 NOT implementable until infant stage VALIDATED via V3-EXQ-586/591 AND behavioural diversity achieved THROUGH the lock via V3-EXQ-543h PASS). claims.yaml 641 -> 645 claims; YAML validates.
  • INV-074/MECH-333/334 Phase-3 crystallization substrate landed 2026-05-17 (ree-v3 f8b93e3). GatedPolicy.crystallize() freezes head_0/head_1/discriminator + adds a fresh plastic expansion MLP (zero-init last Linear so output bit-identical at the transition instant; forward = frozen_gated(x) + expansion(x.detach())); ResidueField.snapshot_ewc_anchor() + ewc_penalty() write-protect established basins (Kirkpatrick 2017 EWC, not a hard freeze); InfantCurriculumScheduler on_phase3_entry fire-once hook. REEConfig.crystallize_at_phase3 default OFF (bit-identical; 484/484 contracts PASS). Nikishin 2023 plasticity injection. F-error pre-check: MECH-314b/c are forward-model-error-dependent (314c is the canonical Pathak 2017 ICM self-defeat) and decay to ~0 before Phase-3, so xtal arms hold MECH-314 novelty-only; MECH-313 / 314a / 320-primary / dACC are F-robust.
  • ARC-062 GAP-B/C/D landed 2026-05-17 (ree-v3 15ca95e). GAP-B head-input first-action one-hot augmentation (bypasses the E2 world-forward compression that flattened SP-CEM first-action diversity to 0.22% of z_world magnitude before the z_world-only heads); GAP-C discriminator-output -> SD-033a rule_state source vector (agent.py gated_policy block reordered before lateral_pfc); GAP-D SD-033a rule_bias_head made optionally trainable. All three bit-identical OFF.
  • ARC-065 SP-CEM flipped to main-path default 2026-05-17 – 6 HippocampalConfig + REEConfig.from_dims defaults flipped (use_support_preserving_cem / support_preserving_stratified_elites / support_preserving_ao_std_floor). INTENTIONAL non-no-op default change (the one deliberate departure from the implement-substrate no-op rule – the legacy collapsing CEM produced the monostrategy that left SD-029 / ARC-062 Rung 2 / goal_pipeline GAP-2/4 / self_attribution GAP-1/2/3 non_contributory). Bit-identical legacy opt-out by explicitly pinning the flags. V3-EXQ-583 3-arm default-wiring equivalence PASS 2026-05-17T09:25Z (ARM_default == ARM_explicit_on within 1e-9, both » ARM_explicit_off). claims.yaml ARC-065 implementation_note; NOT promoted (Rung-1 matched-entropy governance gated on V3-EXQ-569).
  • SD-012 sustained-drive amendment 2026-05-17 – Option 1 drive_ema_alpha trace (EMA of drive_level instead of instantaneous, default 1.0 = bit-identical) + Option 2 drive_floor insatiability floor, both for goal_pipeline:GAP-3. V3-EXQ-582 SD-012 EMA-sweep FAIL (substrate-ceiling: n_contacts_post_warmup=0 all 12 runs) escalated to the V3-EXQ-582a drive_floor sweep.
  • commitment_closure GAP-3 + GAP-11 2026-05-17 – CausalGridWorldV2 env extensions primitives 1-3 (adaptive tolerance-band completion / counter-evidence injection / dual-cue; env-only, bit-identical OFF, 434/434 contracts) + committed_mode_curriculum.py harness helper (P0 warmup / P1 consolidation / P2 eval; mid-probe abort gate). Unblocks the OCD battery V3-EXQ-460b/463b/464b/466b/467b/468b.
• Runner activity since the 2026-05-17T01:11Z nightly: central evidence/experiments/runner_status.json reports 718 -> 720 cumulative completions (+2); all-time breakdown 163 -> 164 PASS (+1) / 285 -> 286 FAIL (+1) / 77 ERROR / 193 UNKNOWN (unchanged); last_updated 2026-05-17T13:11:02Z – the central file is now ~12h stale, the multi-machine runners write per-machine runner_status/<hostname>.json. Cross-machine aggregate as of this read: DLAPTOP-4.local 591 + ree-cloud-1 237 + ree-cloud-2 170 + EWIN-PC 77 + Daniel-PC 28 = 1103 cumulative (274 PASS / 486 FAIL / 107 ERROR / 236 UNKNOWN). The two new central entries: V3-EXQ-582 FAIL (SD-012 EMA-sweep substrate-ceiling) + V3-EXQ-583 PASS (SP-CEM main-path default-wiring equivalence); V3-EXQ-543e/g FAIL routed to the INV-074/MECH-334 crystallization successor.
• Queue depth: 0 -> 11 items. infant_substrate GAP / EXQ-ISEF cohort: V3-EXQ-584 (GAP-7 traj_cosine), 586 (GAP-9 curriculum scheduler), 587 (EXQ-ISEF-001 harm-gradient curriculum, claimed ree-cloud-4), 588 (ISEF-002 transient-benefit z_goal seeding, claimed ree-cloud-2), 589 (ISEF-003 microhabitat latent diversity, claimed DLAPTOP-4.local), 590 (ISEF-004 novelty Goldilocks, claimed ree-cloud-3), 591 (ISEF-005 curriculum-vs-flat) + V3-EXQ-481b (MECH-090 V_s commit-release), 582a (goal_pipeline:GAP-3 Option 2 drive_floor sweep), 592 (commitment_closure:GAP-11 committed-mode curriculum pilot). Cloud capacity scaled: ree-worker-3 (CX43, hel1) + ree-worker-4 provisioned via cloud-scaler.
• Pending review: 0 -> 0 items (pending_review.md regenerated 2026-05-17T12:59:48Z; last review 2026-05-17T10:24:31Z). The day’s failure autopsies (V3-EXQ-572/573 substrate_ceiling; 539/540a-f MECH-307; 541/541a/541b MECH-204; 543f training-signal-gap; 566/577/582 GAP-3) closed all runs into review_tracker.json with no claim weighting changes.
• Bottleneck note (unchanged framing): the ARC-065 behavioural-diversity developmental warm-start failure remains the dominant scientific blocker, now under active experimental test rather than diagnosis. EXQ-ISEF-001 (V3-EXQ-587 harm-gradient curriculum) is Rank 1, claimed by ree-cloud-4 and running – it establishes the warm-start curriculum DEV-NEED-029 requires before MECH-314a / MECH-320 diversity biases are non-zero at cold start, and gates ISEF-002..005 + the deferred Q-043/044/045 + INV-049 retests. The freshly-landed INV-074/MECH-334 crystallization substrate has its falsifier V3-EXQ-543h (2x2x2 use_gated_policy x use_dacc x crystallize_at_phase3, supersedes 543g) queued at priority 10 on ree-cloud-4 – it is the load-bearing test of whether developmental crystallization is the missing piece for ARC-062 behavioural diversity through the lock.

Immediate Work Queue (This Cycle, 2026-05-18)

1. V3-EXQ-587 EXQ-ISEF-001 outcome watch (Rank 1) – harm-gradient curriculum, claimed ree-cloud-4. PASS closes infant_substrate:GAP-10, unblocks ARC-013 / DEV-NEED-004 and the rest of the ISEF-001..005 cohort + the deferred Q-043/044/045 + INV-049 retests.
2. V3-EXQ-543h outcome watch – ARC-062 GAP-B x crystallize_at_phase3 2x2x2 falsifier (INV-074/MECH-334; supersedes 543g; priority 10, claimed ree-cloud-4). Load-bearing test of developmental crystallization as the ARC-062 behavioural-diversity-through-the-lock mechanism.
3. ISEF cohort outcome watch – V3-EXQ-588 (transient-benefit z_goal seeding, ree-cloud-2), 589 (microhabitat latent diversity, DLAPTOP-4.local), 590 (novelty Goldilocks weight sweep, ree-cloud-3), 591 (curriculum-vs-flat baselines).
4. goal_pipeline:GAP-3 – V3-EXQ-582a drive_floor insatiability sweep (Option 2 escalation from the V3-EXQ-582 substrate-ceiling FAIL); PASS closes GAP-3.
5. commitment_closure:GAP-11 – V3-EXQ-592 committed-mode curriculum pilot (EMERGENT / FORCED_RV / STARVED arms); PASS unblocks the OCD battery V3-EXQ-460b/463b/464b/466b/467b/468b.
6. Governance walk – next governance / morning-digest session should absorb the plasticity-crystallization cluster registrations (INV-074/075, MECH-333/334/335/336, ARC-075/076, Q-052) into the recommendation queue and re-confirm 0 pending. ARC-076’s HARD ORDERING GATE means its cluster cannot be promoted until the infant-stage + diversity-through-lock prerequisites validate.

---

Status Snapshot (2026-05-17T01:11Z – nightly docs sync; sleep_substrate GAP-6/GAP-8 complete + GAP-3 unified master flag; infant_substrate GAP-2/3/5/6 env substrates; ARC-067/068 child-MECH design complete (MECH-330 + MECH-331; ARC-068 collapses into MECH-320); calibration-debt diversity sprint diagnoses ARC-065 developmental warm-start failure; governance Proposal G2 backward-traceability checker wired; pending_review 0)

• SDs / MECHs / substrate landings since the 2026-05-13T01:10Z nightly:
  • sleep_substrate GAP-6 + GAP-8 2026-05-15 – StepHarness write-path audit (e1_input scaled by anchor_weight in run_sws_schema_pass) + MECH-272 anchor-channel consumer (mean_anchor threaded through SleepLoopManager._run_cycle -> run_sleep_cycle). V3-EXQ-565 GAP-8 routing-consumer full-runner PASS 2026-05-15T18:03Z (arm0_applied_mean=1.0, arm1~0.6, C1/C2/C3 all True). sleep_substrate_plan.md GAP-6/GAP-8 open -> done; Rung 5 cleared in calibration_debt_index.md.
  • sleep_substrate GAP-3 2026-05-16 – unified REEConfig.use_sleep_aggregation_cluster field + enable_sleep_aggregation_cluster() resolving the eight Phase A-E sub-flags consistently from post_init + end of from_dims (mirrors use_mech307_conjunction / enable_goal_stream). OR-only; MECH-204 + anchor-set / e2_harm_s prereqs deliberately NOT bundled. V3-EXQ-581 owner-EXQ dry-run 6/6 PASS (C1-C5 all four phases fire end-to-end under one flag; C6 ARM_CLUSTER==ARM_EXPLICIT proves pure ergonomics). The 2026-05-16 GAP-4-entry GAP-8/GAP-3 conflation corrected (V3-EXQ-565 is GAP-8’s owner-EXQ).
  • infant_substrate GAP-2/3/5/6 2026-05-16 – CausalGridWorldV2 env substrates: microhabitat Voronoi zones (V3-EXQ-577 FAIL, C2 zone_map_coverage diagnosed as a test-design false-negative – per-episode missing_012 over-constraint vs stochastic Voronoi+ecotone design; substrate functionally validated by C1/C3/C4; V3-EXQ-577a corrected-C2 routed), transient-benefit patches for z_goal seeding (V3-EXQ-578 PASS), pos/zone telemetry (GAP-5), residue-coverage telemetry (ResidueField.get_coverage_telemetry; V3-EXQ-580 PASS 3/3 seeds, ARM_1 harm-gradient coverage 0.77-0.94 » ARM_0 binary 0.23). claims.yaml not modified (governed by full infant pipeline).
  • ARC-067 / ARC-068 child-MECH design 2026-05-16 – two-child split for ARC-067: MECH-330 idle_aversion_acute_restlessness_accumulator (engagement-rate EWMA -> z_harm_a write; 3-input estimator commit-transitions / deliberation-depth / residue-write-rate; Wilson 2014 + Danckert 2018 anchors) + MECH-331 idle_aversion_chronic_anhedonic_flatness_substrate (frontostriatal effort-allocation integrator seeded by sustained MECH-330 non-discharge; apathy archetype per Husain-Roiser 2018; preserves MECH-295 hedonic experience). ARC-068 collapses into MECH-320 per the ARC-068 lit-pull R3 verdict (Niv 2007 mathematical symmetry – the MECH-320 w_passive term IS the ARC-068 implementation). 635 claims; biology-before-formal-definitions gate now fully clear for the non_deficit_action_drives family.
  • governance Proposal G2 2026-05-16 – scripts/check_backward_traceability.py (PyYAML + regex; checks developmental-keyword claims against developmental_needs_register.md Claim IDs; exits 1 on WARNING, --warn-only exits 0; SKIP_TRACEABILITY=1 bypass) wired into governance.sh Step 4b after build_claims_json.py as a hard gate. Current gap: 141 developmental claims, 115 untraced (register coverage ongoing).
• Diagnostic / experimental results (calibration-debt diversity sprint, 2026-05-16):
  • V3-EXQ-569 FAIL – all arms entropy ~0.496, zero diversity lift.
  • V3-EXQ-570 PASS – E2 is not the bottleneck (rollout ratio 52.1).
  • V3-EXQ-571 PASS – F (forward-model) term dominates 88-89% of E3 temporal variance; ALL MECH-313/314/320 + dACC / lateral_pfc / ofc / gated_policy / mech295 diversity-bias components contribute ~0. e3_selector.py + agent.py instrumented with default-OFF score-decomposition flags (349/350 contracts; the one failure pre-existing + unrelated).
  • V3-EXQ-573 NULL – ARC-065 bias-scale 5-10x sweep, all 10 arms bit-for-bit identical.
  • Replay analysis register updates: DEV-NEED-030 (stage-aware replay scheduling) + DEV-NEED-031 (MECH-124 prevention gate); 6 EXP-IDEV-001..006 proposals (317 total); INV-049 evidence_quality_note updated with the infant-content-poverty / waking-consolidation-failure / SWR-immaturity substrate diagnosis (not a theory failure).
• Runner activity since the 2026-05-13T01:10Z nightly: central evidence/experiments/runner_status.json reports 681 -> 718 cumulative completions (+37); all-time breakdown 134 -> 163 PASS / 271 -> 285 FAIL / 76 -> 77 ERROR / 193 UNKNOWN (unchanged); last_updated 2026-05-17T01:11:14Z.
• Queue depth: 0 -> 0 items (empty items: []). The 2026-05-15/16 substrate + governance waves drained the queue; the active goal_pipeline:GAP-3 SD-012 sustained-drive EMA session (TASK_CLAIMS gap3-sustained-drive-ema-20260517T004727Z) holds the next substrate slot.
• Pending review: 0 -> 0 items (pending_review.md regenerated 2026-05-16T20:52:59Z; last review 2026-05-16T18:22:30Z). The 2026-05-15T18:55Z governance cycle walked 5 indexed + 1 runner-only (EXQ-563/563a/563b/565 + 564 cleared; supersession cleanup applied 563->563a->563b), applied 9 hold_pending_v3_substrate (ARC-066/067/068, ARC-070/071, MECH-320, Q-043/044/045), confirmed 0 indexed pending; substrate_queue.json timestamp updated (48 ready, 32 blocked, 38 implemented).
• Bottleneck note (shift): the ARC-065 behavioural-diversity developmental warm-start failure is now the dominant scientific blocker, superseding the pure V_s-monostrategy framing. V3-EXQ-573’s bit-for-bit-identical 10-arm 5-10x bias-scale sweep diagnosed the MECH-314a / MECH-320 diversity biases as literally zero at cold start (not miscalibrated – MECH-313 temperature is irrelevant on near-uniform random-network scores). The right response is a developmental warm-start curriculum (DEV-NEED-029 ARC-065 warm-start gate, PROPOSED), not more bias-scale sweeps. docs/architecture/developmental_experiment_priorities.md (created 2026-05-16) synthesises 7 lit-pull topics + the FAIL/PASS pattern + DEV-NEED register and ranks 10 experiments: EXQ-ISEF-001 (harm-gradient curriculum) is Rank 1, gating everything downstream. Q-043/044/045 and INV-049 retests must be deferred until the warm-start gate is established via EXQ-ISEF-001..006.

Immediate Work Queue (This Cycle, 2026-05-17)

1. goal_pipeline:GAP-3 SD-012 sustained-drive EMA – active session (TASK_CLAIMS gap3-sustained-drive-ema-20260517T004727Z): Option-1 sustained-drive EMA in goal.py + config.py + discriminative alpha-sweep EXQ via /queue-experiment + contract test. Plan-of-record goal_pipeline_plan.md GAP-3.
2. EXQ-ISEF-001 harm-gradient curriculum (Rank 1) – the developmental-experiment-priorities gate. Establishes the warm-start curriculum that DEV-NEED-029 requires before MECH-314a / MECH-320 diversity biases are non-zero at cold start. Gates EXQ-ISEF-002..006 and the deferred Q-043/044/045 + INV-049 retests.
3. V3-EXQ-577a outcome watch – infant_substrate GAP-2 corrected-C2 microhabitat validation (the 577 C2 false-negative redraw guard + corrected acceptance per the failure autopsy).
4. V3-EXQ-581 full-runner watch – sleep_substrate GAP-3 unified-master-flag end-to-end validation (dry-run 6/6 PASS; runner DLAPTOP-4.local auto-claimed).
5. Governance walk – next governance / morning-digest session should re-confirm 0 pending and absorb the calibration-debt sprint results (EXQ-569/570/571 + EXQ-573 NULL) into the ARC-065 / MECH-313/314/320 evidence_quality_notes if not already done.

---

Status Snapshot (2026-05-13T01:10Z – nightly docs sync; MECH-307 default-value recalibration landed (min_drive_to_fire 0.1 -> 0.01 + conjunction_z_beta_threshold 0.6 -> 0.3) after V3-EXQ-540d read-site probe PASS diagnosed substrate-ceiling drive-floor / z_beta-ceiling pattern; V3-EXQ-540e MECH-307 default-fix validation queued and dry-run PASS with first non-zero conj_fire_rate=0.155; V3-EXQ-461 EXP-0157 delayed-reward persistence PASS closes commitment_closure_plan.md GAP-2 at substrate-readiness level; V_s-monostrategy diagnostic cohort queued (V3-EXQ-555/556/557/558 agent-init basin localisation + V3-EXQ-554a decoder/elite-refit localization); runner_remote_control _rrc NameError fix landed; pending_review 0 -> 11 stale-index)

• SDs / MECHs / substrate fixes landed since the 2026-05-12T01:10Z nightly:
  • MECH-307 default-value recalibration 2026-05-12 – two bridge config defaults lowered after V3-EXQ-540d (re-queue of V3-EXQ-540c killed by the 2026-05-12T06:10Z systemctl restart) confirmed at 10x scale (1087 bridge calls / 34784 candidate-reads) that the V3-EXQ-540a / 540b conj_fire_rate=0 cohort was caused by two config defaults sitting above the achievable substrate ceiling under standard env config. mech295_min_drive_to_fire 0.1 -> 0.01 (drive_level max=0.030 / mean=0.016 / frac>0.1=0.000 across 1087 reads – legacy floor never crossed); mech307_conjunction_z_beta_threshold 0.6 -> 0.3 (z_beta_arousal max=0.545 / mean=0.518 – legacy floor above achievable ceiling). Half-tier predicate components (0.3 / 0.15 / 0.3) clear 94.66% of candidates if both gates allowed it through – substrate writes populate read sites cleanly. Changes landed across 4+3 declaration sites (bridge dataclass + REEConfig dataclass + REEConfig.from_dims kwarg + REEAgent getattr fallback). Two contract test assertions updated to match new defaults (test_mech_295_liking_bridge.py default check + test_mech307_consumer_conjunction.py default check + the C4 low-z_beta-blocks scenario uses z_beta_arousal=0.1 instead of 0.4 to remain below the new 0.3 default). 314/314 contracts + 7/7 preflight PASS with new defaults. Backward compat: callers that explicitly set either flag are unaffected. Deferred follow-on (separate session): Option-b semantic fix at mech295_liking_bridge.py:343 (currently reads v[:, 3] legacy unsigned-magnitude rather than v[:, 4] VALENCE_POSITIVE_SURPRISE under Option-b semantics; design-doc fidelity bug, not a behavioural blocker).
  • V3-EXQ-461 EXP-0157 delayed-reward persistence substrate-readiness PASS 2026-05-12T18:18Z on DLAPTOP-4.local closing commitment_closure_plan.md GAP-2 at substrate-readiness level (full behavioural delayed-reward arm remains blocked on GAP-3 CausalGridWorldV2 env extensions). Evidence supports MECH-090 + SD-033a + SD-034 at substrate-readiness level: baseline Hold delay suppression / release, weakened passthrough contrast, SD-033a / MECH-261 replay-gated persistence, strengthened Hold threshold, sd_033a mode-gate table, and SD-034 terminal closure release.
  • runner_remote_control _rrc NameError fix landed 2026-05-12T06:14Z (ree-v3 commit b46c89c). Hoisted runner_remote_control import to module top-level so the _push_remote_heartbeat closure inside experiment_runner.run_experiment resolves _rrc via module globals rather than the main() local scope (closure ran on a heartbeat thread called from a module-level function, raising NameError: name '_rrc' is not defined every tick on both Hetzner cloud workers). Likely introduced by commit ccaabee (heartbeat carries progress + recent_lines for remote progress bars). ree-cloud-1 (91.98.130.117) + ree-cloud-2 (116.203.216.181) systemd services restarted clean via SSH; preflight 7/7 PASS on each; both runners back to idle main-loop polling with no NameError post-restart.
• Strategic note: the 2026-05-12 wave is dominated by two convergent investigations on the V_s-monostrategy substrate ceiling. (a) MECH-307 consumer-conjunction-read recovery: V3-EXQ-540b threshold sweep FAIL -> V3-EXQ-540d read-site probe PASS diagnoses substrate-ceiling drive-floor + z_beta-ceiling -> default-value recalibration -> V3-EXQ-540e behavioural validation queued (PASS expected; would clear goal_pipeline:GAP-1). (b) V_s monostrategy basin geometry localisation in agent-init space: V3-EXQ-552 forced-exploration warmup discovered that seed=7 sustains action_class_entropy~0.68 while seeds 42 / 17 collapse to 0.0 at identical code; V3-EXQ-555 2x2 (env_seed, agent_seed) factorisation queued + smoke-PASSed factorisation invariants; V3-EXQ-557 30-cell agent-seed sweep queued; V3-EXQ-556 8-arm module-init swap diagnostic queued (per-submodule seeded init monkey-patch – though smoke surfaced a replication-failure pattern at the smoke-depth boot loop, prompting V3-EXQ-558 clean seed-pair readout/rank diagnostic without the monkey-patch); V3-EXQ-554a decoder + elite-refit collapse localization re-queued after the 06:10Z restart. The two threads share the same monostrategy floor and inform one another – if 540e PASSes, the consumer-side blocker is downstream of (and possibly independent of) the seed-anomaly basin.
• Runner activity since the 2026-05-12T01:10Z nightly: central evidence/experiments/runner_status.json reports 674 -> 681 cumulative completions (+7). All-time breakdown: 136 -> 134 PASS (-2) / 270 -> 271 FAIL (+1) / 75 -> 76 ERROR (+1) / 193 -> 193 UNKNOWN (+0). V3 subset 652 runs (125 PASS / 258 FAIL / 76 ERROR / 193 UNKNOWN). The negative PASS delta is reconciliation indexing (some prior PASS entries reclassified per evidence_direction_per_claim review; manifest content unchanged). New entries since last nightly: V3-EXQ-540c (ERROR, SIGTERM at restart), V3-EXQ-540d (PASS, MECH-307 read-site probe), V3-EXQ-461 (PASS, EXP-0157 delayed-reward persistence). The other ~4 entries are indexer-side surfacing of earlier completions newly included after the post-2026-05-11 governance walk.
• Queue depth: 1 -> 0 items (empty items: []). Yesterday’s V3-EXQ-540b (MECH-307 consumer-conjunction threshold sweep) completed FAIL; today’s wave (V3-EXQ-554a / 555 / 556 / 557 / 558 monostrategy diagnostics + V3-EXQ-540e MECH-307 default-fix validation) were all pre-claimed by runners before the nightly read so the central queue file is empty at this snapshot. substrate_queue.json unchanged from yesterday (the day’s substrate landing was the MECH-307 default-value recalibration, which is a config tweak rather than a new substrate so does not surface a new entry).
• Pending review: 0 -> 11 items (1 PASS / 6 FAIL / 4 runner-only). The pending_review.md regenerated 2026-05-12T18:15Z surfaced V3-EXQ-549 / 550 / 543d / 540a / 540b / 540e FAILs plus V3-EXQ-540c MECH-307 readsite probe PASS (which is the manifest the V3-EXQ-540d re-queue produced – same script name, runner output bookkeeping displays the queue_id but the indexer reads the script’s output_file). Runner-only stale-index items: V3-EXQ-552 (FAIL), V3-EXQ-555 (PASS), V3-EXQ-557 (PASS), V3-EXQ-540c (ERROR). All eleven are governance-walk eligible but accumulated because no governance walk has fired today; the next morning-digest / governance run will drain them.
• Plan-of-record progressions (2026-05-12): commitment_closure_plan.md GAP-2 status open -> done at substrate-readiness level (V3-EXQ-461 EXP-0157 PASS); full behavioural arm is gated on GAP-3 CausalGridWorldV2 env extensions. goal_pipeline_plan.md GAP-1 v3_pending behavioural-validation gate is still v3_pending pending V3-EXQ-540e outcome (PASS clears it; the 2026-05-11 GAP-1 status open -> done covered the substrate landing not the behavioural validation).
• Bottleneck note: the V_s-monostrategy substrate ceiling continues to be the dominant scientific bottleneck. Three concurrent threads share that floor (V_s basin geometry localisation via V3-EXQ-555 / 556 / 557 / 558; MECH-307 consumer-side recalibration via V3-EXQ-540e; ARC-062 Phase 3 wiring via V3-EXQ-543d). The V_s basin geometry thread is now the load-bearing diagnostic – if a single submodule’s seed-7 init is sufficient to escape monostrategy (R1 result on V3-EXQ-556), that localises the substrate-ceiling cause to one specific weight initialisation distribution; if it’s conjunctive (R2), every submodule’s init contributes; if it’s no individual module (R3-R5), the agent-side seed effect is more diffuse than per-module init differences alone explain.

Immediate Work Queue (This Cycle, 2026-05-13)

1. V3-EXQ-540e outcome watch – MECH-307 default-fix behavioural validation queued on DLAPTOP-4.local. Dry-run smoke at 6 ep / 1 seed produced ARM_2_full conj_fire_rate=0.155 (first non-zero conjunction firing since substrate landed). PASS routes to clearing goal_pipeline:GAP-1 v3_pending behavioural-validation gate and unblocking GAP-2 SD-049 Phase 2 V3-EXQ-514 behavioural validation. FAIL routes to a structural read-site audit + the deferred Option-b semantic fix at mech295_liking_bridge.py:343 (v[:, 3] -> v[:, 4]).
2. V3-EXQ-555 outcome watch – 2x2 (env_seed, agent_seed) factorisation diagnostic localising whether the V3-EXQ-552 seed-7 entropy anomaly is env-side, agent-side, conjunctive, or either-sufficient. Pre-registered 5-row interpretation grid (R1 env_side_only / R2 agent_side_only / R3 conjunctive / R4 either_sufficient / R5 replication_failure).
3. V3-EXQ-557 outcome watch – 30-cell agent-seed sweep at fixed env_seed=42 bounding the size of the monostrategy basin in agent-init space. Builds on V3-EXQ-555 by characterising how many agent-init seeds escape (R1 deep_collapse_basin >= 80% collapsed; R2 bimodal; R3 mostly_diverse; R4 continuous_spread).
4. V3-EXQ-556 + V3-EXQ-558 outcome watch – 8-arm module-init swap (V3-EXQ-556 via class-level __init__ monkey-patch with per-module seed override; smoke surfaced a replication-failure pattern that V3-EXQ-558 routes around by reusing the V3-EXQ-555 factored-seed helper for a clean seed-pair readout/rank diagnostic instead). Pre-registered 6-row grid: R1 single_module_sufficient / R2 multi_module_conjunctive / R3 latent_stack_only / R4 e3_only / R5 hippocampal_only / R6 replication_failure.
5. V3-EXQ-554a outcome watch – decoder + elite-refit collapse localization (5 measurement points M0-M4 inside HippocampalModule.propose_trajectories via experiment-side monkey-patch). Pre-registered 5-row interpretation grid: R1 decoder_collapse / R2 elite_refit_collapse / R3 noise_shape_irrelevant / R4 e2_rollout_collapse / R5 other.
6. Governance walk for the 11 pending_review items – next morning-digest / governance session should drain V3-EXQ-549 / 550 / 543d / 540a / 540b / 540e FAILs plus V3-EXQ-540c read-site probe PASS plus the four stale-index runner-only items (552 / 555 / 557 / 540c). Re-run python scripts/generate_pending_review.py to confirm 0 pending after the walk.

---

Status Snapshot (2026-05-12T01:10Z – nightly docs sync; MECH-307 Option-b Gap-1 substrate landing closes goal_pipeline:GAP-1; SD-054 bipartite layout extension; MECH-323 + MECH-324 ARC-071 child claims registered; governance cycle applied 16 hold_pending_v3_substrate decisions and reclassified the V_s monostrategy cohort; self_attribution_plan GAP-1 inverted to monostrategy substrate-ceiling; V3-EXQ-550 z_goal monostrategy falsifier FAIL at no-training depth; V3-EXQ-551/551a pipeline-entropy diagnostic localised entropy=0.0 cliff to CEM proposer; pending_review 0 -> 0)

• SDs / MECHs landed / claims registered since the 2026-05-11T01:10Z nightly:
  • MECH-307 Option-b Gap-1 substrate landing 2026-05-11. Per user override 2026-05-11 (option-b over the design-doc default option-a), the SD-014 valence vector is extended from 4 to 6 components with new constants VALENCE_POSITIVE_SURPRISE=4 and VALENCE_NEGATIVE_SURPRISE=5; MECH-205 PE write site dispatches between three paths (Option-b split / Option-a signed-single-channel / legacy unsigned magnitude) routing surprise by concurrent harm_signal sign while ALSO writing magnitude to legacy VALENCE_SURPRISE for backward-compat. New convenience master use_mech307_conjunction (default False) flips all three substrate-side sub-flags (use_mech307_split_surprise + use_mech307_schema_multichannel + use_mech307_predicted_location_write) via REEConfig.post_init. 309/309 contracts + 7/7 preflight PASS with master OFF. Field-level smoke verified: split-channel writes accumulate correctly under MECH-094 simulation gate. SD-014 6-channel amendment retained as registered fallback. Closes goal_pipeline_plan.md GAP-1; GAP-2 (SD-049 Phase 2 behavioural validation under MECH-307-fixed substrate) unblocked.
  • SD-054 bipartite layout extension 2026-05-11. Three new CausalGridWorldV2.__init__ kwargs (reef_bipartite_layout, reef_bipartite_axis, reef_bipartite_agent_band_radius) that partition reef and food cells into disjoint geometric halves with the agent spawning in a midline band. Forces reef-bound vs forage-bound trajectories to have categorically opposite first-action argmaxes by construction – targets the CEM-candidate distinguishability bottleneck surfaced by V3-EXQ-543b diagnose-errors. V3-EXQ-548 substrate-readiness PASS 2026-05-11 confirmed 1.27x structural divergence uplift (0.633 -> 0.807); legacy mode bit-identical OFF.
  • MECH-323 + MECH-324 ARC-071 child claims registered 2026-05-11. MECH-323 policy.composition.chunk_accumulator_formation (striatum/DLS-analog formation operator – joint AND trigger over repetition count + outcome-variance + evaluative gate with V_s-positive secondary preference; Sutton 1999 options-framework structure; MECH-094-strict default + MECH-322 sleep-replay carve-out). MECH-324 policy.composition.chunk_maintenance_dissolution (IL/vmPFC-analog maintenance operator – four-state lifecycle FORMING / CRYSTALLISED / DISSOLVING / DISSOLVED with crystallisation counter + outcome-variance dissolution gate + replay-origin accelerated-dissolution per MECH-322). ARC-071 substrate now FULLY SPECIFIED at the claim level (formation + maintenance + sleep-replay carve-out); substrate-level implementation in ree_core deferred to a separate /implement-substrate session; first three-arm validation EXQ deferred until substrate landed. claim count 618 -> 619 -> 620.
• Strategic note: the 2026-05-11 wave is dominated by the V_s monostrategy substrate-ceiling thread. Three concurrent diagnostic / falsifier sessions converged on the same underlying issue: (a) V3-EXQ-550 z_goal monostrategy falsifier FAIL at no-training depth (action_class_entropy=0.0 in BOTH arms; z_goal pipeline wired-but-inert with z_goal_update_calls=1200 + z_goal_norm_peak=0.0 across all ARM_ON seeds); (b) self_attribution_plan GAP-1 forensic read inverted from arbitration-data to substrate-ceiling (V3-EXQ-445/445b produced floating-point-identical metrics across architectural arms – bit-identical signature of trajectory determinism under V_s monostrategy); (c) V3-EXQ-551 + V3-EXQ-551a pipeline-entropy diagnostic PASS localised the entropy=0.0 cliff to the CEM proposer stage. V3-EXQ-553 orthogonal CEM seeding is the substrate-side fix under test; V3-EXQ-552 forced-exploration warmup is the parallel training-data-narrowness vs substrate-collapse falsifier. The MECH-307 conjunction architecture lands its Gap-1 substrate the same day, but V3-EXQ-540a’s C2 consumer-conjunction-read FAIL signals the consumer-side thresholds need recalibration (V3-EXQ-540b 4-arm sweep queued and claimed by Mac). ARC-062 Phase 3 wiring continues to be the architectural side gate (V3-EXQ-543c FAIL on probe-gate; V3-EXQ-543d 2x2 factorial of {gated_policy x dACC} with MECH-260 anti-recency at 0.5 queued).
• Runner activity since the 2026-05-11T01:10Z nightly: central evidence/experiments/runner_status.json reports 665 -> 674 cumulative completions (+9). All-time breakdown: 136 PASS (+5) / 270 FAIL (+4) / 75 ERROR (+0) / 193 UNKNOWN (-2). V3 subset 639 runs (116 PASS / 257 FAIL / 73 ERROR / 193 UNKNOWN). PASS deltas (+5): V3-EXQ-548 SD-054 bipartite substrate-readiness, V3-EXQ-551 + V3-EXQ-551a pipeline-entropy diagnostic, V3-EXQ-547 MECH-320 tonic vigor substrate-readiness (carried over from the 2026-05-10 wave indexer-side surfacing). FAIL deltas (+4): V3-EXQ-550 z_goal monostrategy falsifier, V3-EXQ-543c ARC-062 Phase 3 falsifier, V3-EXQ-540a MECH-307 3-arm gap decomposition, V3-EXQ-141c MECH-111 novelty-drive RNG-desync indexer-side surfacing.
• Queue depth: 0 -> 1 item at this nightly read. V3-EXQ-540b (MECH-307 consumer-conjunction threshold sweep, claimed by DLAPTOP-4.local 2026-05-12T01:07Z; 4 arms varying the wanting / liking / z_beta consumer thresholds with substrate fully ON in every arm). Concurrent runs not yet in the queue at this read (queue-then-claim races): V3-EXQ-543d (ARC-062 Phase 3 2x2 factorial), V3-EXQ-549 (ARC-066 / MECH-320 discriminative pair), V3-EXQ-552 (forced-exploration warmup), V3-EXQ-553 (orthogonal CEM seeding).
• Pending review: 0 -> 0 items. The 2026-05-11T20:10Z review-exq550-543c session walked V3-EXQ-550 + V3-EXQ-543c FAILs; both reclassified per pre-registered grids without claim weighting changes (V3-EXQ-550 evidence_direction_per_claim FAIL supports MECH-269 V_s monomodal hold at no-training depth; V3-EXQ-543c non_contributory ARC-062 Phase 3 wiring inert under the SD-054 bipartite substrate before V3-EXQ-543d redesign). Earlier in the day, the 2026-05-11T17:13Z governance cycle walked 3 indexed pending + 5 unclaimed + 2 runner-only pending experiments and applied 16 hold_pending_v3_substrate recommendations covering 5 ARC (ARC-066/067/068/070/071) + 3 Q (Q-043/044/045) + 8 section-level duplicates.
• Plan-of-record progressions (2026-05-11): goal_pipeline_plan.md GAP-1 status open -> done via MECH-307 Option-b substrate landing; self_attribution_plan.md GAP-1 status open -> blocked (monostrategy substrate-ceiling, same blocker as GAP-2); arc_062_rule_apprehension_plan.md GAP-B resume_condition recorded V3-EXQ-543c FAIL diagnosis (probe-gate + ARC-062 architectural-isolation); SD-049-PHASE-3 SD-032 consumer cascade migration remains deferred per goal_pipeline_plan.md GAP-5; substrate_queue.json grew with the new entries reflecting today’s substrate landings + cluster registrations.
• Renderer / tooling improvement: closure tab in serve.py / closure.html now distinguishes resolved-lineage edges from active blockers (faded dotted rendering when depends_on predecessor is done or deferred*; tooltip surfaces blocked on / resume when rows above phase/severity; resume_condition free-text passthrough). CLOSURE_VERSION bumped 2026-05-09.3 -> 2026-05-11.1.
• Bottleneck note: the V_s-monostrategy substrate ceiling is now the dominant scientific bottleneck across at least four open-claim threads (SD-029 self-attribution; SD-032b dACC arbitration; ARC-062 rule apprehension; MECH-307 consumer-conjunction read). V3-EXQ-553 orthogonal CEM seeding is the immediate substrate-side test (proposer-stage entropy cliff). Underneath that, the ARC-070 / ARC-071 R6 SAFETY-CRITICAL governance decision on MECH-094 hypothesis_tag strict-vs-relaxed gating for the chunking write path now has BOTH substrate slots specified (MECH-323 formation + MECH-324 maintenance) – the governance call is the gate before any implement-substrate session.

Immediate Work Queue (This Cycle, 2026-05-12)

1. V3-EXQ-540b outcome watch – MECH-307 consumer-conjunction threshold sweep currently running on DLAPTOP-4.local. PASS routes to a 540c behavioural retest at the chosen threshold; PARTIAL C1-PASS-C2-FAIL routes to a structural read-side audit (suspect: kernel-decay reading nearby zeros, RBF center drift, or write/read site z_world mismatch).
2. V3-EXQ-553 orthogonal CEM seeding outcome watch – proposer-stage substrate-side test of the V_s monostrategy cliff localised by V3-EXQ-551/551a. PASS clears the proposer-as-cliff-variable reading; FAIL routes to downstream stage 2/3 review.
3. V3-EXQ-552 forced-exploration warmup outcome watch – parallel training-data-narrowness vs substrate-collapse falsifier on the V3-EXQ-550 monostrategy finding.
4. V3-EXQ-543d outcome watch – ARC-062 Phase 3 wiring 2x2 factorial of {gated_policy x dACC} with MECH-260 anti-recency=0.5. PASS = cluster wiring is the missing piece (both substrates contribute).
5. ARC-070 / ARC-071 R6 SAFETY-CRITICAL governance decision – dedicated session for the MECH-094 hypothesis_tag strict-vs-relaxed call for chunking writes; both MECH-323 + MECH-324 substrate slots now registered, the governance verdict gates any /implement-substrate session.
6. MECH-323 / MECH-324 substrate implementation – after the R6 verdict lands, the ree_core substrate work for the formation + maintenance operators is the next architectural pass.

---

Status Snapshot (2026-05-11T11:35Z – scheduled PM /lit-pull + /update-docs; Q-045 lit-pull adds 5 entries / lit_conf 0.9 / evidence_quadrant plausible_unproven; R1 verdict COUPLED-NOT-COLLAPSED extends Q-045 resolution categories; MECH-260 anchor flagged missing in arc_065 Pull 1 RECOVERED via Scholl 2015; no SD / MECH status changes)

• Literature additions since the 2026-05-11T01:10Z nightly snapshot:
  • Q-045 dedicated lit-pull – 5 entries landed in evidence/literature/targeted_review_q_045/ covering the LC-NE tonic noise (MECH-313) vs dACC anti-recency (MECH-260) substrate-independence question. Tervo et al. 2014 (Cell, DOI 10.1016/j.cell.2014.08.037) R1 LOAD-BEARING: LC-NE input drives ACC stochastic-mode switching with ACC engagement SUPPRESSED during stochastic mode – the two substrates are coupled at the circuit level, neither fully independent nor fully collapsed. Scholl et al. 2015 (J Neurosci, DOI 10.1523/JNEUROSCI.0396-15.2015) R2 substrate-distinctness CONFIRMED + recovers MECH-260 anchor flagged missing in arc_065 Pull 1 synthesis 2026-05-10. Kennerley et al. 2006 (Nat Neurosci, DOI 10.1038/nn1724) causal lesion complement. Meder et al. 2017 (Nat Commun, DOI 10.1038/s41467-017-02169-w) multi-timescale value spectrum reinforces substrate-distinctness from a third angle. Yu & Dayan 2005 (Neuron, DOI 10.1016/j.neuron.2005.04.026) theoretical anchor.
• Verdicts and design refinements:
  • R1 COUPLED-NOT-COLLAPSED (mixed-direction evidence from Tervo 2014) – MECH-313 and MECH-260 substrate-distinct AND circuit-coupled; resolution categories EXTENDED to include “DIRECTIONALLY COUPLED” as a fourth category alongside the three originally registered (mutually load-bearing / 313-dominant / 260-dominant).
  • R4 LOAD-BEARING design refinement – current Q-045 4-arm ablation is INSUFFICIENT to expose the Tervo asymmetry; needs extension to 8-cell (4-arm x 2-LC-amplitude) OR addition of LC -> ACC coupling ablation. both-ON should NOT be linear superposition of singletons.
  • R5 substrate-readiness precondition – SD-054 reef substrate temporal-horizon must be verified BEFORE 4-arm authorisation; Kennerley-style across-trial integration only emerges with sufficient temporal horizon and a single-tick outcome substrate cannot dissociate MECH-260 from MECH-313 in the Kennerley sense.
• Indexer delta: literature entries 1326 -> 1331 (+5); claim_evidence.v1.json Q-045 entry: literature_confidence 0.0 -> 0.9, overall_confidence 0.0 -> 0.9, evidence_quadrant plausible_unproven, direction_counts 4 supports + 1 mixed.
• Hygiene win: recovers the MECH-260 Scholl 2015 anchor that the ARC-065 Pull 1 synthesis flagged as un-retrievable 2026-05-10 (citation-lookup ambiguity resolved by author + journal + year combination via PubMed author-field search).
• Governance impact: Q-045 evidence_quality_note extended in claims.yaml with R1-R5 verdicts + reference to targeted_review_q_045/synthesis.md. NO claim status changes (Q-045 remains open; MECH-313 remains candidate_substrate_landed; MECH-260 remains substrate_landed). validate_claims –strict OK 68 invariants; build_claims_json wrote 618 claims.
• Bottleneck note (unchanged from 2026-05-11T01:10Z): ARC-062 Phase 3 wiring pass + ARC-064/ARC-065/ARC-066 child-MECH cluster behavioural validation remain the immediate gate. Q-045 4-arm ablation is now better-scaffolded but still gated on the Phase 3 redesign + the SD-054 substrate-readiness precondition R5 surfaces.

Immediate Work Queue (This Cycle, 2026-05-11 – PM addendum to nightly)

1. Carry-forward from nightly: all 6 items unchanged (ARC-062 Phase 3 wiring redesign + Q-043/044/045 behavioural ablation cohort design + ARC-070/071 R6 SAFETY-CRITICAL governance + MECH-321 substrate + ARC-067/068 child-MECH design + V3-EXQ-141c diagnose-errors).
2. Q-045 design extension follow-on – before queueing the Q-045 4-arm ablation, decide whether to extend it to the 8-cell (4-arm x 2-LC-amplitude) form per R4 verdict OR add a dedicated LC -> ACC coupling ablation arm. Sequence the SD-054 multi-trial-outcome-dependency substrate-readiness diagnostic (R5 precondition) BEFORE the 4-arm authorisation.

---

Status Snapshot (2026-05-11T01:10Z – nightly docs sync; four ARC-064/ARC-065/ARC-066 child substrates land (MECH-313 noise floor + MECH-314 structured curiosity + MECH-319 simulation-mode rule-write gate + MECH-320 tonic vigor coupling); two new architectural cluster registrations (ARC-066/067/068 non_deficit_action_drives + ARC-069/070/071 policy_primitive_granularity); MECH-163 depends_on +ARC-071 per ARC-071 lit-pull R3 verdict; ARC-062 Phase 3 falsifier V3-EXQ-543b FAIL on Mac; pending_review 2 -> 0; runner_status.json single-file deprecated – per-machine aggregate now reported)

• SDs / MECHs moved to Implemented since the 2026-05-10T01:10Z nightly snapshot:
  • MECH-313 (ARC-065 child) stochastic noise floor (LC-NE tonic / SAC analog) IMPLEMENTED 2026-05-10. Module ree_core/policy/noise_floor.py (NoiseFloor + NoiseFloorConfig). Pure-arithmetic regulator (no learned parameters); sibling to MECH-314 / MECH-318 / MECH-319 in the ree_core.policy package. Single primitive noise_floor.compute_effective_temperature(baseline_temperature, simulation_mode) returns max(baseline_T + noise_floor_alpha, noise_floor_min_temperature) – SAC-entropy-bonus analog (Haarnoja 2018) on E3 softmax temperature. State-independent; complement to MECH-104 phasic spike. Distinct from MECH-260 dACC anti-recency (state-dependent); Q-045 falsifies whether they collapse. Phase-1 instantiation choice = SEPARATE module at the e3.select() call site (revisit at Q-045 4-arm ablation). MECH-094: simulation_mode=True returns baseline temperature unchanged + increments skip counter only. V3-EXQ-544 substrate-readiness 5/5 PASS UC1-UC5 smoke (manifest PASS; runner outcome flag ERROR per the substrate-readiness false-ERROR stdout sentinel pattern, fixed mid-day via the diagnose-errors session that landed verdict: print across 542/544/545). 11 contract tests in tests/contracts/test_mech_313_noise_floor.py PASS.
  • MECH-314 (ARC-065 child) structured curiosity bonus + 3 sub-flavours (314a/b/c) IMPLEMENTED 2026-05-10. Module ree_core/policy/structured_curiosity.py (StructuredCuriosity + StructuredCuriosityConfig). Pure-arithmetic, no learned parameters; sibling to MECH-313 NoiseFloor. Three sub-flavours implemented as a single module with master + 3 independently-togglable sub-switches per Pull 1 R3 verdict: 314a striatal novelty (per-candidate min-distance from candidate’s first-step z_world to nearest active ResidueField RBF center, normalised by candidate-pool mean norm); 314b frontopolar uncertainty (e3._running_variance scalar, broadcast across [K] in Phase 1); 314c learning progress (EMA of |PE_t - PE_{t-K}|, broadcast scalar in Phase 1). Composed AFTER MECH-295 liking-bridge block and BEFORE MECH-313 noise_floor temperature lift (orthogonal axes). Phase 1 honest-scoping caveat: 314b/c are state-dependent broadcast scalars; per-candidate refinement deferred to Phase 2 follow-on. Q-044’s three-arm ablation IS a flag-set decision – the substrate guarantees each sub-flavour can be turned on/off independently. MECH-094: simulation_mode=True returns zeros[K] + increments skip counter only. V3-EXQ-545 substrate-readiness 5/5 PASS UC1-UC5 (ran twice on Mac + cloud-2 via multi-machine race). 13 contract tests in tests/contracts/test_mech_314_curiosity.py PASS; 273/273 contracts + 7/7 preflight PASS with master OFF.
  • MECH-319 (arc_062 GAP-K) simulation-mode rule-write gate IMPLEMENTED 2026-05-10. Module ree_core/regulators/simulation_mode_rule_gate.py (SimulationModeRuleGate + SimulationModeRuleGateConfig + SimulationModeRuleGateDiagnostics). Substrate-level instantiation of MECH-094 at the rule-arbitration layer per Pull 3 SYNTHESIS R1 GENUINE-NOVELTY-CONFIRMED + Pull 4 R3 KEEP-AS-IS verdicts. Pure-arithmetic regulator (sibling to GABAergicDecayRegulator and BroadcastOverrideRegulator). Single primitive gate.effective_simulation_mode(simulation_mode, site) -> bool with truth-table semantics: master OFF identity passthrough; master ON + admit_writes=False blocks sim writes (MECH-319 normal); master ON + admit_writes=True (V3-EXQ-543c-successor falsifier control) admits sim writes. Two existing arbitration-write call sites in REEAgent.select_action() consult the gate when instantiated: GatedPolicy block (replace literal simulation_mode=False with gate.effective_simulation_mode call) + LateralPFCAnalog block (skip update() when blocked, compute_bias still runs). Per-site diagnostic counters on {gated_policy, lateral_pfc, default}. Construction raises ValueError on admit_writes=True without master ON (loud-not-silent guard). MECH-094 NOT modified per KEEP-AS-IS verdict. V3-EXQ-546 substrate-readiness 6/6 PASS UC1-UC5 + UC3b precondition (ran twice on Mac + cloud-2 via multi-machine race); 15 contract tests in tests/contracts/test_mech_319_simulation_mode_rule_gate.py PASS; 288/288 contract + preflight tests PASS with master OFF. claims.yaml MECH-319 candidate -> candidate_substrate_landed; v3_pending: true retained pending V3-EXQ-543c-successor falsifier with admit_writes=True. arc_062 GAP-K closed (registered -> substrate_landed).

  • MECH-320 (ARC-066 child) tonic vigor coupling score bias (mesolimbic-DA-vigor analog) IMPLEMENTED 2026-05-10. Module ree_core/policy/tonic_vigor.py (TonicVigor + TonicVigorConfig + TonicVigorOutput). First child mechanism for ARC-066 (the non_deficit_action_drives architectural family). Pure-arithmetic regulator (sister to MECH-313 NoiseFloor + MECH-314 StructuredCuriosity in ree_core.policy). Composed AFTER MECH-314 curiosity (orthogonal axis at the candidate-feature level) and BEFORE MECH-313 noise_floor (orthogonal regulator at the temperature level). Algorithm: v_t = max(0, slow EWMA over realised E3 score) * gate_energy * gate_drive * gate_pe; bias[i] = -w_action*v_t on action classes / +w_passive*v_t on noop class (additive primary; multiplicative gain falsifiable secondary via tonic_vigor_form=”multiplicative” – distinguishable on a held-out non-uniform-

    score

    batch). TARGET-FREE: bias applies regardless of whether any z_goal is currently active – closes the “well-fed-safe-familiar agent has no positive gradient to act” gap that ARC-066 registered. Defaults: half_life=100 (long-window per R4 verdict), w_action=w_passive=0.1, bias_scale=0.1, gate_energy_min=0.2, gate_drive_max=0.7, gate_pe_max=1.0, noop_class=0. MECH-094: simulation_mode=True on either compute_score_bias or update_score_receipt returns zeros + increments skip counter only. V3-EXQ-547 substrate-readiness 6/6 PASS UC1-UC6 on cloud-2 2026-05-10T20:56Z; 28 contract tests in tests/contracts/test_mech_320_tonic_vigor.py PASS; 309/309 contracts + 7/7 preflight PASS with master OFF. claims.yaml MECH-320 candidate -> candidate_substrate_landed.

  • ARC-066 / ARC-067 / ARC-068 cluster registration (non_deficit_action_drives family) 2026-05-10. Three architectural-slot claims registered as the family principle that behaviour comes from surplus capacity AND from deficits, not deficits alone: ARC-066 tonic_vigor_coupling (capacity -> action bias); ARC-067 idle_aversion_boredom (sustained low-engagement is aversive); ARC-068 opportunity_cost_no_op_penalty (waiting carries cost). Three companion lit-pulls landed: ARC-066 lit_conf 0.789 supports (LC-NE substrate REJECTED, mesolimbic DA-vigor LOAD-BEARING per Niv 2007 + Salamone & Correa 2012 + Beierholm 2013); ARC-067 lit_conf 0.85 supports; ARC-068 lit_conf 0.806 supports-direction-dominant (R1 SEPARATE-AT-ARCHITECTURE-VIA-KERNEL not via substrate; R3 ARC-066 + ARC-068 collapse LICENSED at implementation layer per Niv 2007 mathematical symmetry but slot-level separation preserved for psychiatric failure-mode dissociation). Anchored in docs/architecture/non_deficit_action_drives.md. MECH-320 (ARC-066 first child) substrate landed same day; ARC-067 / ARC-068 child-MECH design pending.
  • ARC-069 / ARC-070 / ARC-071 cluster registration (policy_primitive_granularity family) 2026-05-10. Three architectural-slot claims registered as the family principle that the unit of policy operated on is itself dynamic (not fixed): ARC-069 parent (policy_hierarchy_dynamic_regranularisation); ARC-070 decomposition-on-prediction-failure (zoom in / re-segment when an imagined chunk fails to ground); ARC-071 composition-via-repeated-grounding (zoom out / chunking). Two companion lit-pulls landed: ARC-070 lit_conf 0.88 supports (R2 LOAD-BEARING SHARED SUBSTRATE – ARC-070 implemented as bidirectional extension of MECH-288 event_segmenter, NOT a new module; Schacter 2008 constructive-episodic-simulation core network supplies empirical anchor); ARC-071 lit_conf 0.848 supports (R3 LOAD-BEARING – CONFIRMED ARC-071 IS the missing transition mechanism in MECH-163 dual_goal_directed_systems, MECH-163 depends_on extended +ARC-071 the same day; R6 SAFETY-CRITICAL escalation – biology does NOT cleanly gate chunking write path against replay/imagined sequences per Albouy 2013, ARC-071’s pre-registered MECH-094 hypothesis_tag=False strict-gating MORE CONSERVATIVE than biology, governance decision pending). Anchored in docs/architecture/policy_primitive_granularity.md. MECH-321 (ARC-070 first child mechanism, policy.decomposition_via_event_segmenter) registered candidate / v3_pending the same day with depends_on ARC-070 + MECH-288 + MECH-269 + MECH-094.
• Strategic note: today landed four child substrates in three distinct architectural clusters (ARC-064 GAP-K MECH-319 + ARC-065 MECH-313/314 + ARC-066 MECH-320), plus two new architectural-slot cluster registrations (ARC-066/067/068 non_deficit_action_drives + ARC-069/070/071 policy_primitive_granularity), plus MECH-321 as ARC-070 first child. The substrate-readiness diagnostics for all four child substrates passed today; behavioural validation of the ARC-066 + ARC-065 cohorts (Q-043 / Q-044 / Q-045 cross-cohort ablation) is the next-up scientific work and will exercise the V3-EXQ-543c-successor admit_writes=True falsifier substrate. ARC-062 Phase 3 wiring (gated_policy bias-head into E3 optimizer + discriminator -> SD-033a LateralPFCAnalog.update() source vector) was attempted in V3-EXQ-543b on Mac (~4h, FAIL); commitment_closure_plan.md GAP-1 remains blocked pending Phase 3 design refinement.
• Runner activity since the 2026-05-10T01:10Z nightly: the 665 figure cited at the prior nightly was a single-machine read (Mac local). The single-file runner_status.json no longer exists – per the multi-machine coordination policy each machine writes to evidence/experiments/runner_status/<hostname>.json. Per-machine aggregate as of 2026-05-11T01:10Z: 555 DLAPTOP-4.local + 28 Daniel-PC + 77 EWIN-PC + 222 ree-cloud-1 + 163 ree-cloud-2 = 1045 cumulative. Cumulative cross-machine breakdown: 241 PASS / 463 FAIL / 105 ERROR / 236 UNKNOWN. +10 cumulative-by-machine completions today (deduped: 8 unique experiments because V3-EXQ-545 + V3-EXQ-546 each ran twice via the multi-machine claim race): V3-EXQ-543b (ARC-062 Phase 3 falsifier on Mac, FAIL ~4h); V3-EXQ-544 (MECH-313 noise floor) substrate-readiness “false ERROR” with manifest PASS; V3-EXQ-545 (MECH-314 structured curiosity) PASS x2 on Mac + cloud-2; V3-EXQ-546 (MECH-319 simulation-mode rule-write gate) PASS x2 on Mac + cloud-2; V3-EXQ-547 (MECH-320 tonic vigor coupling) PASS on cloud-2; V3-EXQ-514h + V3-EXQ-514i (SD-049 Phase 2 behavioural successors) PASS on cloud-1; V3-EXQ-141c (MECH-111 novelty-drive RNG-desync) FAIL on cloud-2.
• Queue depth: 0 -> 0 items (empty items: []) at this nightly read; today’s queue churn (V3-EXQ-543b + 544/545/546/547 + 514h/i + 141c) all completed during the day. substrate_queue.json grew 69 -> 78 entries with the new ARC-064 / ARC-065 / MECH-313 / MECH-314 / MECH-316 / MECH-317 / MECH-318 / MECH-319 / SD-054 entries reflecting today’s substrate landings + cluster registrations. ARC-062 entry status updated candidate_v3_pending -> phase_1_implemented.
• Pending review: 2 -> 0 items. The 2026-05-10T12:24Z governance cycle walked the 3 PASS pending experiments (V3-EXQ-500a SD-017 readiness probe + V3-EXQ-543 ARC-062 Phase 2 falsifier already-non_contributory + V3-EXQ-503a SD-017 sleep discriminative pair) and applied 13 pending_user recommendations via decision_log.v1.jsonl appends (10 hold_pending_v3_substrate on ARC-062, ARC-064, ARC-065, MECH-309, MECH-312, MECH-313, MECH-316, MECH-317, MECH-318, SD-054 newly registered v3_pending claims; 3 hold_candidate_resolve_conflict on ARC-045, MECH-166, MECH-204). The 2026-05-10T12:18Z diagnose-errors session also added the four held ERRORs (V3-EXQ-538 / 495 / 449c / 455a) to discussed_experiment_dirs with explicit rationale.
• Plan-of-record progressions (2026-05-10): arc_062_rule_apprehension_plan.md GAP-K closed (MECH-319 substrate landed); GAP-I MECH-318 absorption check VERDICT B partial (no new V3 substrate; empirical retire-vs-promote deferred to V3-EXQ-543c-successor on multi-rule-context substrate); 14 lit entries added across two ARC-066/067/068 lit-pulls + two ARC-070/071 lit-pulls + one ARC-068 lit-pull = ~30 lit entries lifting cluster lit_conf 0.0 -> 0.8+ across all 6 newly registered architectural-slot claims. 2 new umbrella architecture docs (non_deficit_action_drives.md + policy_primitive_granularity.md).
• Bottleneck note: the ARC-062 Phase 3 wiring pass + ARC-064/ARC-065/ARC-066 child-MECH cluster behavioural validation is the immediate gate. Phase 3 wiring was attempted in V3-EXQ-543b on Mac (~4h, FAIL); design needs refinement before another attempt; commitment_closure_plan.md GAP-1 remains blocked. The four ARC-064/ARC-065/ARC-066 child substrates all landed substrate-readiness PASS today; the next move is the Q-043 / Q-044 / Q-045 cross-cohort behavioural ablation. Underneath those, the ARC-070/071 R6 SAFETY-CRITICAL governance decision (MECH-094 hypothesis_tag strict-vs-relaxed for chunking write path, escalated by the ARC-071 lit-pull synthesis) and the non_deficit_action_drives + policy_primitive_granularity child-MECH design cycles are the architectural-side gates that flow from today’s two cluster registrations. The monostrategy / reef-recovery thread continues to be the underlying scientific bottleneck pending ARC-062 Phase 3 + the rule-apprehension cluster’s behavioural validation cohort.

Immediate Work Queue (This Cycle, 2026-05-11)

1. ARC-062 Phase 3 wiring redesign – V3-EXQ-543b FAILed at ~4h; design needs refinement before another attempt. Closes commitment_closure_plan.md GAP-1.
2. Q-043 / Q-044 / Q-045 behavioural ablation cohort design – now that all four ARC-064/ARC-065/ARC-066 child substrates have landed (MECH-313 / MECH-314 / MECH-319 / MECH-320), queue the cross-cohort relative-weight calibration (Q-043) + sub-flavour independence (Q-044) + MECH-313 vs MECH-260 collapse falsifier (Q-045) on the V3-EXQ-543c-successor admit_writes=True falsifier substrate.
3. ARC-070/071 R6 SAFETY-CRITICAL governance decision – dedicated session for the MECH-094 hypothesis_tag strict-vs-relaxed decision for the chunking write path. Touches MECH-094 / ARC-071 / MECH-292/293 / sleep_substrate / SD-039. ARC-071 child-MECH design is gated on this verdict.
4. MECH-321 substrate work – after MECH-288 event_segmenter.py input_stream label extension lands, MECH-321 (ARC-070 first child mechanism) substrate is a thin policy-side BoundaryEvent consumer per the ARC-070 R2 bidirectional-substrate verdict.
5. ARC-067 / ARC-068 child-MECH design – cycle through the same design-doc + lit-pull-driven cycle as ARC-066 / MECH-320 used today.
6. V3-EXQ-141c diagnose-errors – pending_review=0 at nightly close but the 141c FAIL on cloud-2 is the immediate /diagnose-errors candidate per the bit-identical-arms measurement-validity pattern; it shares the V3-EXQ-141d governance-deferred status but landed on a different machine.

---

Status Snapshot (2026-05-10T01:10Z – nightly docs sync; ARC-062 Phase 1 lands gated-policy heads + context discriminator; Phase 2 GAP-B monomodal-collapse falsifier PASS; MECH-204 step-size sweep cohort PASS; sleep_substrate GAP-2 Tier-1 cohort 3 PASS + 2 FAIL drains the queue; pending_review 1 -> 2)

• SDs / MECHs moved to Implemented since the 2026-05-09T01:10Z nightly snapshot:
  • ARC-062 Phase 1 (gated-policy heads + context discriminator) IMPLEMENTED 2026-05-09 – arc_062_rule_apprehension_plan.md GAP-A. New ree_core/policy/gated_policy.py (GatedPolicy + GatedPolicyConfig + GatedPolicyOutput): N=2 scoring heads sharing E3 candidate features (symmetry-broken init on heads’ last-Linear bias) + 3-stream context discriminator on (z_world, z_self, z_harm_a) per Pull A SYNTHESIS R1 (multi-stream input) at score_bias level per R3; n_heads=2 substrate-constrained per R2 (raises ValueError otherwise). disc_init_scale=0.1 keeps sigmoid output near 0.5 at init; bias clamped to [-bias_scale, +bias_scale]. NO connection to SD-033a in Phase 1 – that wiring is Phase 3 (closes commitment_closure_plan.md GAP-1). REEAgent.select_action composes gated_policy_score_bias additively into dacc_score_bias before the MECH-295 block. MECH-094 simulation_mode=True returns zeros and increments only the skip counter. 5 contract tests tests/contracts/test_gated_policy.py (C1-C5); full ree-v3 suite 249/249 PASS. Substrate-readiness V3-EXQ-542 5/5 manifest PASS UC1-UC5 on Mac 2026-05-09T20:22Z (runner outcome flag ERROR with manifest verdict PASS per the substrate-readiness pattern).

  • ARC-062 Phase 2 GAP-B (monomodal-collapse falsifier on SD-054 reef) PASS 2026-05-09 – V3-EXQ-543 on Mac in ~50min: ARM_0 use_gated_policy=False vs ARM_1c use_gated_policy=True with full 3-stream discriminator at ARM_1_med density (hazard_food_attraction=0.7, n_reef_patches=3, reef_patch_radius=2, n_hazards=4, size=12). PASS rule >=2 of {C2 state-dependence (Spearman binned drive-vs-reef abs >= 0.20 AND ARM_1c

    rho

    > ARM_0

    rho

    ), C3 risk-type dissociation (forage_hazard_rate / transit_hazard_rate >= 0.50 relative magnitude), C4 cross-seed variation (CoV(reef_visit_fraction) >= 0.10)} with no F1 (total invariance / monomodal-collapse signature) or F2 (biologically inverted: rho > +0.40 monotone refuge-use under chronic high-drive). Phase 3 wiring pass (discriminator -> SD-033a LateralPFCAnalog.update() source vector + adds bias-head parameters to E3 optimiser) is the next-up consumer; closes commitment_closure_plan.md GAP-1.

  • MECH-204 step-size sweep cohort PASS 2026-05-09 – V3-EXQ-541a / 541b / 541c on DLAPTOP-4.local / Mac. Companion EXP-0171 step-size sweep gated on V3-EXQ-541 PASS surfaced as a 3-arm cohort step values {0.05 / 0.10 / 0.25 / 0.50}; 541c PASS at 16-cycles (the wider-cycle dose-response sweep) with ARM_4 step=0.5 clearing the C4 5% threshold at 9.03% and ARM_3 step=0.25 at 4.51% (high-end defensible; supported the default bump from 0.1 -> 0.25 in the 2026-05-09T13:36Z sleep-substrate-phase1-closure session).

  • Sleep_substrate_plan.md GAP-2 Tier-1 successor cohort completed 2026-05-09 – 5 EXQs queued in series (265a, 500a, 503a, 436a, 418l) with the validated Phase 2 substrate template (sd016_writepath_mode=’off’, sd016_diversification_weight=0.5, use_per_stream_vs=True, use_anchor_sets=True, use_sd039_anchor_payload=True). Outcomes: V3-EXQ-265a PASS 20:12Z (sleep-on signed

    diff

    > 0.05 in >=2/3 seeds across SWS / SWS_THEN_REM contrasts), V3-EXQ-500a PASS 20:41Z (single-arm SD-017 sleep-phase readiness probe), V3-EXQ-503a PASS 21:46Z (FULL_4_PHASE_ON vs NO_SLEEP_BASELINE discriminative pair with C4 cross-arm

    M2 diff

    > 0.20 calibrated from –dry-run smoke), V3-EXQ-436a FAIL 21:52Z (3 conditions x 5 seeds multi-claim [‘SD-017’, ‘ARC-045’, ‘MECH-166’] with per-seed distribution diagnostics; failure routes to /diagnose-errors per-seed-distribution-grid not substrate retraction), V3-EXQ-418l FAIL 21:53Z (SD-016 action_bias_div retest single-claim [‘SD-017’]; failure routes to /diagnose-errors). 4 of 5 successors clear; sleep_substrate_plan.md GAP-2 row update happens after governance walks the failures.

• Strategic note: the 2026-05-09 wave drained the queue completely. The empty queue is the natural inflection point for governance to walk the cohort and dispatch GAP-2 row update + ARC-062 Phase 3 wiring + the two FAIL diagnostic root-cause cycles before any new behavioural EXQs are queued.
• Runner activity since the 2026-05-09T01:10Z nightly: 654 -> 665 completions (+11). Cumulative breakdown 124 -> 131 PASS / 263 -> 266 FAIL / 72 -> 73 ERROR / 195 -> 195 UNKNOWN. PASS deltas (+7): V3-EXQ-541a/541b/541c MECH-204 step-size cohort, V3-EXQ-265a sleep-on signed-diff PASS, V3-EXQ-500a SD-017 readiness probe PASS, V3-EXQ-503a discriminative pair PASS, V3-EXQ-543 ARC-062 Phase 2 falsifier PASS. FAIL deltas (+3): V3-EXQ-436a multi-claim sleep cohort, V3-EXQ-418l SD-016 action_bias_div retest, plus one carryover FAIL surfaced in the indexer. ERROR delta (+1): V3-EXQ-542 (manifest verdict PASS but runner outcome ERROR per substrate-readiness pattern).
• Queue depth: 1 -> 0 items (empty items: []). V3-EXQ-540 (MECH-307 3-arm gap decomposition, queued 2026-05-08, claimed by Mac 2026-05-09T00:00Z) is no longer in the queue at this read – presumed run earlier in the 2026-05-09 wave; verify against runner_status. Next-up substrate work in evidence/planning/substrate_queue.json remains SD-049 Phase 2 z_resource encoder follow-on (ready=True priority=2) and the ARC-062 Phase 3 wiring pass that closes commitment_closure_plan.md GAP-1 (now unblocked by the V3-EXQ-543 Phase 2 PASS).
• Pending review: 1 -> 2 items. V3-EXQ-530c (ARC-016 precision-commit StepHarness retest, carry-over from 2026-05-08T22:34Z governance, deferred to /diagnose-errors per the bit-identical-arms measurement-validity pattern) plus V3-EXQ-141d (MECH-111 novelty-drive RNG-desync, also FAIL deferred to /diagnose-errors). Both are deferred not awaiting a governance walk.
• Plan-of-record progressions (2026-05-09): sleep_substrate_plan.md Phase 1 closure landed (GAP-1 status open -> done; default rem_precision_recalibration_step bumped 0.1 -> 0.25 in REEConfig backed by V3-EXQ-541c PASS; Phase 7 reframed deferred-conditional -> deferred-to-V4-unless-future-evidence-reverses); Phase 2 (SD-017 retest cohort) Tier-0 status correction blocked -> open (EXQ-418e A2_div_only had cleared slot_diversity 0.5 threshold cleanly back on 2026-04-27); arc_062_rule_apprehension_plan.md registered as a sibling plan-of-record + GAP-A done + GAP-B queued; commitment_closure_plan.md GAP-1 reframed open -> blocked on arc_062_rule_apprehension:GAP-A/B; substrate_queue.json ARC-062 entry added at queue tail (priority 2, ready=true); SD-033 governance plan-doc bookkeeping closure (CHK-PUSH open -> done; closure-graph 87.5% -> 100%); 14 lit entries added across two ARC-062 lit-pulls (8 entries on lateral PFC rule-context modulation + 6 entries on refugia-vs-forage behavioural ecology) lifting cluster lit_conf 0.0 -> 0.890 across MECH-309 / ARC-062 / ARC-063 with SD-054 picking up 6 direct-tagged entries.
• Bottleneck note: the ARC-062 cluster (MECH-309 / ARC-062 / ARC-063 / SD-054 rule-apprehension cluster registered 2026-05-08) is now the governance-side gate. Phase 3 wiring (discriminator -> SD-033a + bias-head into E3 optimiser) is the architectural next step; commitment_closure GAP-1 unblock cascades through it. Underneath, the monostrategy / reef-recovery thread is no longer the dominant scientific bottleneck – the rule-apprehension cluster has displaced it as the load-bearing diagnostic for whether REE V3 needs a non-Bayesian rule-creator architectural slot or whether the V3 weak reading (gated-policy heads) is sufficient. The two FAILs in the sleep_substrate GAP-2 Tier-1 cohort (V3-EXQ-436a / 418l) are the only items on the governance-walk side that need /diagnose-errors per-seed-distribution diagnostics before the GAP-2 row can flip done.

Immediate Work Queue (This Cycle, 2026-05-10)

1. Sleep_substrate_plan.md GAP-2 row update – 4 of 5 Tier-1 successors PASS (265a / 500a / 503a + the prior 265a anchor); the row can flip in-progress -> done once the 436a / 418l FAILs are walked through /diagnose-errors and either superseded or accepted as-tagged.
2. ARC-062 Phase 3 wiring pass – discriminator output threaded into SD-033a LateralPFCAnalog.update() source vector + adds gated_policy bias-head parameters to E3 optimiser. Closes commitment_closure_plan.md GAP-1. Phase 2 PASS (V3-EXQ-543) clears the gate.
3. V3-EXQ-530c + V3-EXQ-141d /diagnose-errors – two pending_review FAILs both deferred to root-cause diagnosis under the bit-identical-arms measurement-validity pattern.
4. V3-EXQ-540 MECH-307 outcome – if the run completed during the 2026-05-09 wave, governance walks the C1/C2/C3 acceptance grid + ARM_3 SD-014 6-channel fallback decision. If runner_status shows it surfaced UNKNOWN, the manifest needs review per the substrate-readiness pattern.
5. Next-substrate decision – with the queue empty and SD-049 Phase 2 follow-on + ARC-062 Phase 3 + V3-EXQ-495 V3-full-completion-gate all eligible, governance picks the load-bearing next move.

---

Status Snapshot (2026-05-09T01:10Z – nightly docs sync; sleep-substrate Phase 1 lands MECH-204 recalibration consumer; MECH-307 conjunction architecture in V3-EXQ-540 3-arm gap decomposition; governance cycle applies MECH-307 hold + Q-040 narrow-open with three sub-questions; pending_review 0 -> 1)

• SDs / MECHs moved to Implemented since the 2026-05-08T01:11Z nightly snapshot:
  • MECH-204 sleep substrate Phase 1 IMPLEMENTED 2026-05-08 – sleep_substrate_plan.md GAP-1. SerotoninModule.compute_recalibration_target() exposes the captured precision_at_rem_entry zero-point reference; E3TrajectorySelector.recalibrate_precision_to(target, step) applies Option A linear interpolation new_rv = (1-step)*rv + step*(1.0/(target+1e-6)); WRITEBACK-phase sibling step in SleepLoopManager._run_cycle runs independently of MECH-273 self-model gradient. New REEConfig fields use_rem_precision_recalibration (default False; bit-identical OFF) + rem_precision_recalibration_step (default 0.1 per plan-of-record Q1). Contract suite test_mech204_precision_recalibration.py 9/9 PASS covering C1 surface, C2 default-OFF, C3 sleep_loop-ON / recalibration-OFF no-metrics, C4 arithmetic, C5/C6 zero-target / zero-step no-op, C7 capture-only regression guard, C8 WRITEBACK firing end-to-end, C9 drift movement. Full preflight + contracts 237/237 PASS. V3-EXQ-541 validation ran 2026-05-08T23:43Z with runner-level FAIL outcome and result_summary “verdict: PASS” – governance walk reconciliation pending.
  • MECH-307 anticipatory affect conjunction architecture (consumer side) IMPLEMENTED 2026-05-08 – MECH295LikingBridge.compute_conjunction_score_bias() reads SD-014 valence + z_beta arousal at per-candidate predicted-imminent locations and applies a negative approach bias when the four-way conjunction holds. New REEConfig flag use_mech307_consumer_conjunction_read (default False; bit-identical OFF). Wired into REEAgent.select_action() additively with the legacy m295_bias. test_mech307_consumer_conjunction.py 8/8 PASS; full contract suite 221/221 + 7/7 preflight PASS.
  • Rule-apprehension cluster registration 2026-05-08 – MECH-309 (monomodal-collapse-as-equilibrium-without-rule-apprehender, mechanism_hypothesis candidate); ARC-062 (V3 weak reading, implementation_phase=v3, v3_pending=true); ARC-063 (V4 strong reading, implementation_phase=v4); SD-054 reef enrichment substrate (renamed from SD-050 to disambiguate; v3_pending flipped to true). docs/thoughts/2026-05-04_Waking_rule_apprehension_later_sleep_schema.md promoted to docs/architecture/rule_apprehension_layer.md as canonical anchor. EXP-0171 three-arm discriminative proposal queued.
  • Plan-of-record registration wave (2026-05-08) – four new evidence/planning/*_plan.md docs registered as resume primitives across multi-session work: sleep_substrate_plan.md (8-gap inventory + 7-phase sequencing for SD-017 / MECH-204 / sleep cluster), commitment_closure_plan.md (10-gap inventory + 8-phase sequencing for SD-034 / MECH-090/091 / MECH-260 / MECH-266/267/268 / SD-033a/b cluster, with sd033_governance_plan.md retained as OCD-specific test-battery sub-plan), self_attribution_plan.md (5-gap inventory + 5-phase sequencing for SD-029 / MECH-256 / MECH-257 / SD-013 / ARC-033 / ARC-058 / MECH-258 / MECH-260 + V4 SD-030/SD-031), goal_pipeline_plan.md (6-gap inventory + 6-phase sequencing for SD-012 + SD-014 + SD-015 + SD-018 + SD-049 + MECH-216 + MECH-229/230 + MECH-117/295/307 + ARC-030/032/036/051). Substrate_queue back-fills: MECH-267 + MECH-268 + SD-018 (queue 61 -> 68 entries).
• Strategic shift (2026-05-08): sleep substrate flagged as the top-priority unblocker. Per user direction, “we may not make progress elsewhere until progress is made there.” Sleep_substrate_plan.md Phase 1 now landed; Phase 2 (SD-017 retest cohort gated on EXQ-418e SD-016 div-loss A2/A3 arms) and Phase 7 (deferred-conditional MECH-204 Option B broadcast read at action-selection time) remain ahead. The plan converges with goal_pipeline_plan at the SD-049 sleep-on cohort boundary (V3-EXQ-514 family with use_sleep_loop / sws_enabled / rem_enabled flags).
• Runner activity since the 2026-05-08T01:11Z nightly: 631 -> 654 completions (+23). Cumulative breakdown 120 -> 124 PASS / 255 -> 263 FAIL / 72 -> 72 ERROR / 184 -> 195 UNKNOWN. Notable PASS: V3-EXQ-244a stale-ERROR -> PASS reconciliation (manifest at runs/v3_exq_244a…/manifest.json was already PASS / supports MECH-165; runner SIGTERM’d on ree-cloud-2 after writing manifest but before bookkeeping success); V3-EXQ-514g StepHarness wider-seed sweep PASS on ree-cloud-1; V3-EXQ-106a SD-011 harm_obs_a temporal persistence re-validation PASS (3.2s on ree-cloud-1). Notable FAIL: V3-EXQ-539 MECH-307 4-arm substrate-readiness (substrate counters PASS but C5 behavioural FAIL – legacy MECH-295 cue path does not read the conjunction signal; led to V3-EXQ-540 3-arm gap decomposition with consumer-conjunction-read ON in all arms); V3-EXQ-541 MECH-204 Phase 1 validation (FAIL outcome flag; in-script verdict PASS – governance walk reconciliation pending); V3-EXQ-526 Q-034 reef threshold sweep; V3-EXQ-454a ARC-016 adaptive commitment under reef.
• Queue depth: 7 -> 1 item. Only V3-EXQ-540 (MECH-307 3-arm gap decomposition + Path B consumer conjunction read; ARM_0_off / ARM_1_signed_pe / ARM_2_full; priority=5; machine_affinity=DLAPTOP-4.local; 70 episodes x 3 seeds x 3 conditions; estimated 90 min) is currently queued, claimed by Mac at 2026-05-09T00:00:27Z. ARM_3 SD-014 6-channel fallback DEFERRED.
• Pending review: 0 -> 1 item. The 2026-05-08T22:34Z governance cycle walked 10 indexed pending FAILs (4 superseded predecessors EXQ-537/537b/537c/141c flipped to superseded; 3 already-triaged accepted as-tagged; 530c held for /diagnose-errors per bit-identical-arms measurement-validity pattern; 537d + 539 accepted as-tagged). 9 run_ids added to reviewed_run_ids. The 2026-05-08T22:38Z pending_review.md regeneration shows V3-EXQ-530c as the only remaining pending item.
• Governance decisions applied 2026-05-08T22:34Z (2 pending_user resolved):
  • MECH-307 hold_pending_v3_substrate decision logged; v3_pending=true; V3-EXQ-540 is the queued discriminative validation. evidence_quality_note appended to MECH-307 in claims.yaml capturing the EXQ-539 read + consumer-conjunction-read landing + EXQ-540 queued chain.

  • Q-040 narrow_open_question applied – decomposed into Q-040.a (factorial 2x2 of {MECH-269b OFF/ON} x {SD-032b OFF/ON} on EXQ-483a retest, load-bearing arbitration), Q-040.b (alternative-hypothesis isolator: MECH-295 liking-stream wiring? cross-witness EXQ-536b force-arm probe), Q-040.c (mechanism quantification: dACC weight delta proportional to

    precision-weighted forward-PE

    ? EXQ-475a-conditions retest). Empirical resolution gated on StepHarness migration + MECH-307 substrate landing + MECH-269b V_s-rollout-gate.

• Manifest reclassifications (2026-05-08): 4 flips (537/537b/537c/141c -> superseded) with evidence_direction_per_claim mapped to superseded for each tagged claim, evidence_direction_note appended documenting the lettered successor (537d for the 537 chain, 141d for 141c). 11 manifests in the prior fix-update-z-goal-bug cohort patched with [update_z_goal_typeerror_swallowed] note (4 superseded, 7 trace-only); evidence_quality_note added to 8 affected claims (Q-040, MECH-269b, MECH-295, SD-037, MECH-280, MECH-281, SD-036, MECH-279).
• Substrate fixes (2026-05-08): ree-v3/runner_remote_control.py active-claim guard broadened from evidence/experiments/ substring to evidence/ substring after a 2026-05-08T~18:25Z silent autostash revert of an evidence/planning/substrate_queue.json edit (matching the 2026-04-29 EXQ-232 evidence/experiments/ incident signature). Contract test_active_claim_evidence_guard.py 8/8 PASS. ree-v3/experiment_runner.py REE_QUEUE_ID + REE_RUNNER_SIGNAL_DIR env-leak fix: child subprocess no longer inherits stale parent shell env when queue_id is falsy or signal_dir is None; 6-test contract test_runner_env_isolation.py PASS.
• Bottleneck note: the sleep-substrate Phase 1 validation + MECH-307 conjunction-architecture cluster is the immediate gate. V3-EXQ-541 outcome reconciliation (FAIL flag with verdict PASS) is the first governance task; whether step=0.1 default is calibrated correctly for the C1/C2/C3 acceptance grid or whether EXP-0171 step-size sweep needs to fire ahead of the C5 behavioural arm. V3-EXQ-540 outcome (in flight on Mac) decides whether the four-gap conjunction architecture is the right fix or whether the SD-014 6-channel fallback (ARM_3, deferred) becomes the natural follow-on. Underneath those, the monostrategy / reef-recovery thread remains the dominant scientific bottleneck, now reframed by the rule-apprehension cluster registration (MECH-309 / ARC-062 / ARC-063) – whether the missing primitive is a V3 gated-policy architectural slot or a V4 distributed CandidateRule field with tolerance gate.

Immediate Work Queue (This Cycle, 2026-05-09)

1. V3-EXQ-540 outcome watch (MECH-307 3-arm gap decomposition, Mac, ~90 min). C1 substrate-readiness counter dissociation, C2 conjunction-fire rate >= 0.10 in ARM_2 only, C3 approach_commit_rate ARM_2 >= ARM_0 + 0.10 in 2/3 seeds AND ARM_2 > ARM_1. Outcome dispatches whether four-gap conjunction architecture is the right fix or whether SD-014 6-channel fallback (ARM_3) becomes the natural follow-on.
2. V3-EXQ-541 MECH-204 Phase 1 governance walk reconciliation – runner-level FAIL outcome with manifest result_summary “verdict: PASS” needs reconciliation. Step=0.1 calibration check; potential EXP-0171 step-size sweep gate before the C5 behavioural arm.
3. V3-EXQ-530c /diagnose-errors (single pending_review item; bit-identical-arms measurement-validity pattern, deferred from 2026-05-08 governance cycle for diagnostic root-cause).
4. Plan-of-record Phase progression – sleep_substrate_plan Phase 2 (SD-017 retest cohort gated on EXQ-418e SD-016 div-loss A2/A3 arms producing slot_diversity >= 0.5), commitment_closure_plan Phase 1 (SD-033a bias-head training), self_attribution_plan Phase 1 (V3-EXQ-445h forensic read; not gated on substrate work), goal_pipeline_plan Phase 1 (already in progress via MECH-307 / V3-EXQ-540).
5. Rule-apprehension cluster activation – EXP-0171 three-arm discriminative proposal (ARM_0 baseline / ARM_1 reef-only / ARM_2 reef + gated-policy stub) gated on V3-EXQ-540 outcome.
6. StepHarness audit of governance write paths (shared concern across sleep_substrate_plan GAP-6 and commitment_closure_plan GAP-10 – combine when either reaches its respective phase).

---

Status Snapshot (2026-05-08T01:11Z – nightly docs sync, quiet runner day; bug-fix retest cohort kicked off; MECH-295 goal-seeding diagnostic queued; pending_review back to 0; Q-041 registered)

• No new SD / MECH structural landings since the 2026-05-07T01:12Z nightly snapshot. The 2026-05-07/08 sessions were governance / triage / bug-fix-retest construction, not architectural substrate work. Two concrete substrates landed in the experimental tooling: ree-v3/experiments/_harness.py StepHarness (canonical sense / update_z_goal / update_residue sequence; kwargs-only call shape; no bare-except wrappers) + ree-v3/experiments/_metrics.py canonical extractors (extract_dacc_score_bias reads the actual [K] tensor passed to E3.select, not the raw bundle slice) – both fix the 2026-05-07T23:30Z fix-update-z-goal-bug cohort’s update_z_goal positional/kwarg TypeError + bare-except silent-swallow class. 189/189 contract tests PASS post-fix.
• Queue depth: 0 -> 7 items (3 claimed + 4 pending). The 2026-05-07T20:35Z bugfix-requeue-433f-483b-476c-490f-445h-514f-523b session wrote and queued seven lettered-iteration corrected scripts fixing two ree_core bugs (Bug 1 BreathOscillator disabled, Bug 2 _committed_step_idx saturation in non-bistable path; 476c also required an agent.act() -> sense()/select_action() API migration). Three are now claimed and running – V3-EXQ-433f (SD-029 reef comparator on ree-cloud-1), V3-EXQ-483b (SD-037 broadcast override 4-arm on ree-cloud-2), V3-EXQ-514f (SD-049 Phase 2 reef behavioural validation on DLAPTOP-4.local). V3-EXQ-523b (SD-029 reef-unblocked comparator) is pending. Two MECH-295 goal-seeding follow-ons from the goal-seeding-diagnostic-followup-2026-05-07T2255Z session are pending: V3-EXQ-536a (per-step instrumentation dispatching H_a drive-collapse-on-contact / H_b threshold never crossed / H_c update fires but z_goal does not grow) and V3-EXQ-536b (z_goal_inject force-arm via MECH-188 hook isolating upstream-vs-downstream blocker). V3-EXQ-537 (SD-029/MECH-256 single-pass residual replacing the V3-EXQ-535a two-pass cf_gap) is pending after the 2026-05-08T00:15Z multi-sense audit reclassified V3-EXQ-535a non_contributory for both SD-029 and MECH-256.
• Runner activity since the 2026-05-07T01:12Z nightly: 628 -> 631 completions (+3). Cumulative breakdown 120 -> 120 PASS / 255 -> 255 FAIL / 72 -> 72 ERROR / 181 -> 184 UNKNOWN. The three new completions are V3-EXQ-535a (SD-029 P3 eval-fix attempt; reclassified non_contributory in the multi-sense audit – script computes two-pass cf_gap not the single-pass residual the spec requires), the V3-EXQ-530b ARC-016 precision-commit rename-rerun (renamed from V3-EXQ-530 to bypass the runner_status.json completed-record skip after the original SIGTERM ERROR; runner claimed it on DLAPTOP-4.local within 21 seconds), and the V3-EXQ-433f / V3-EXQ-483b queue claims that flowed through but are still in flight at this nightly read. All three surface as UNKNOWN result codes pending the next governance walk.
• Pending review: 32 -> 0 items. The active governance cycle (claim governance-2026-05-06T2156Z, opened 2026-05-06T21:56Z) finished walking the 32-item accumulation overnight (status: done per WORKSPACE_STATE.md Recent Work and the 2026-05-08T00:43Z pending_review regeneration showing 0 items). The 2026-05-07T23:00Z rv-pinned-cluster-review session walked the EXQ-514d/514e/524/530/536 substrate-bug cluster (root cause: experiments call select_action() but never agent.update_residue(), so post_action_update() / running_variance update never fires; precision pinned at 1/0.5; current_precision=1.999996 constant). The 2026-05-08T00:15Z triage-7-weighting-multi-sense audit cleared 7 weighting scripts after the multi-sense bug uncovered: SUPERSEDED EXQ-124 (env.reset mid-episode + double-sense weakening MECH-033), EXQ-143 (eval-time double-sense asymmetric across hopfield ON/OFF arms weakening MECH-118), EXQ-490f (cohort bug); trace-noted (direction kept) EXQ-108 / EXQ-375 / EXQ-047i/j/k – evidence_quality_note added to MECH-033, MECH-118, MECH-135, MECH-073, MECH-095, SD-005. 983 runs indexed.
• Q-041 registered (2026-05-07T22:25Z): register-threshold-supervisor session formalised the unified meta-level threshold supervisor research direction (default-toward-Q approach: scattered adaptive loci ARC-016 / SD-032c/d/e / MECH-040 / MECH-204 inventoried; supervisor SD/MECH cluster deferred until evidence directs); docs/architecture/threshold_supervisor_survey.md written with timescale-spread table from per-step EMA (~10 steps) up to pACC drive_bias (~347-step half-life) plus three architectural gaps (cross-substrate volatility tracking, MECH-204 sleep-writeback absent, joint setpoint drift); EXP-0170 added to manual_proposals.v1.json as low-priority exploratory_probe testing cross-substrate threshold coherence under sustained drive_level=0.9 vs 0.5 baseline (PCA-based primary analysis, gated on V_s-monostrategy substrate clearing). claims 580 -> 581, proposals 266 -> 267, 68 invariants validated.
• Bottleneck note: the bug-fix retest cohort + MECH-295 goal-seeding diagnostic is the immediate gate. Whether the BreathOscillator + saturating-step-idx fix recovers behavioural acceptance under SD-029 / SD-037 / SD-049 Phase 2 will be answered by V3-EXQ-433f / 483b / 514f as they land. Underneath that, the EXQ-536 z_goal_active_fraction=0.0 finding (“goal seeding is upstream-blocked OR commit chain is inert even with seeded z_goal”) is the load-bearing diagnostic queued ahead – V3-EXQ-536a / 536b will dispatch root cause for the MECH-295 wired-but-inert pattern that survived V3-EXQ-490c + V3-EXQ-490e FAILs. The monostrategy / reef-recovery thread (whether reef substrate alone is sufficient under SD-029 / SD-032b / SD-032c / Q-034 / MECH-112 / ARC-016 / MECH-257 / SD-049 Phase 2 or whether downstream calibration is required) remains the dominant underlying scientific bottleneck.

Immediate Work Queue (This Cycle, 2026-05-08)

1. Bug-fix retest cohort runtime watch – V3-EXQ-433f (SD-029 reef comparator, ree-cloud-1, ~250 min), V3-EXQ-483b (SD-037 broadcast override 4-arm, ree-cloud-2, ~250 min), V3-EXQ-514f (SD-049 Phase 2 reef behavioural validation, DLAPTOP-4.local, ~100 min) running concurrently. Outcome dispatches whether the BreathOscillator + saturating-step-idx fix is the recovery lever for the monostrategy-blocked predecessors.
2. V3-EXQ-536a / 536b MECH-295 goal-seeding diagnostic – whichever runner picks these up first dispatches root cause for the EXQ-536 z_goal_active_fraction=0.0 finding (H_a drive-collapse on contact, H_b benefit threshold never crossed, H_c update fires but z_goal does not grow; vs upstream seeding versus downstream commit chain).

3. V3-EXQ-537 SD-029/MECH-256 single-pass residual – supersedes V3-EXQ-535a. Replaces the two-pass cf_gap with the SD-029 / MECH-256 spec single-pass residual

   z_obs - E2(z_prev, a_actual)

   . ARM_0 baseline dropped (no ext events). ARM_1 intact + ARM_2 action-scrambled falsification. P2 graduation raised to 0.85 with interventional training (SD-013) enabled. C1 forward_r2 threshold relaxed to 0.85 to accommodate reef_food_ext-induced noise.

4. rv-pinned cluster architectural follow-up – the 2026-05-07T23:00Z review surfaced an architectural hazard: experiments that call select_action() but never agent.update_residue() leave the BG gate silently dormant (precision pinned, running_variance never updates). Either document the update_residue() loop contract explicitly or move rv update into select_action() so the gate cannot be silently dormant; consider Q-claim.
5. EWIN-PC + ree-cloud-1/2 SIGTERM coincidence – the 2026-05-06T09:28-09:29Z three-EXQ kill pattern noted in the previous nightly recurs in the form of a runner_status ERROR record blocking V3-EXQ-530 force_rerun (had to rename to V3-EXQ-530b to bypass). Supervisor / restart strategy on the cloud workers warrants follow-up.
6. Active governance cycle close-out – claim governance-2026-05-06T2156Z is now done per the 2026-05-08T00:43Z pending_review=0 regeneration; final claims.yaml / claims.json deltas land in the next nightly snapshot.

---

Status Snapshot (2026-05-07T01:12Z – nightly docs sync, 2026-05-06 in-flight wave drained the queue + 32-item pending_review accumulation under active governance walk)

• No new SD / MECH structural landings since the 2026-05-06T01:35Z nightly snapshot. All architectural substrate work captured in the previous nightly (SD-019a / SD-050-comparator / SD-050 reef substrate / SD-051 / SD-052 / SD-049 Phase 2 hybrid encoder) remains current. Today’s repo activity was experimental: the 2026-05-06 morning’s “write backlog EXQ-528-534” pass + EXQ-244a + EXQ-535 SD-029 P1 target fix + manifest-correction wave (EXQ-452a duplicate marked superseded; EXQ-433c marked superseded; EXQ-433e re-tagged from bespoke inconclusive_insufficient_events to canonical non_contributory) + the EXQ-418k canonical run_id restoration pass (manifests + dir + source scripts 418j/418k all corrected; bare-timestamp run_id phantom eliminated) + the indexer backlog literature-evidence epoch-filter parity fix + the Q-019 disconfirming-balance lit pull (4 weakens entries: Redgrave 1999 unified-selection; Mink 1996 focused selection; GPR 2001 Part I selection+control decomposition; Wouterlood 2018 cross-loop dopaminergic linkage; lit_conf 0.884 -> 0.776; quadrant unchanged plausible_unproven).
• Runner activity since the 2026-05-06T01:35Z nightly: 608 -> 628 completions (+20). Cumulative breakdown 117 -> 120 PASS / 253 -> 255 FAIL / 70 -> 72 ERROR / 168 -> 181 UNKNOWN. Three new PASSes – V3-EXQ-528 SD-029 comparator-trained (the 60-episode trained-comparator clears the SD-029 graduation gate – supersedes the older fixed-2000-step EXQ-523 substrate-readiness baseline), V3-EXQ-533 MECH-102 harm-stream ablation (substrate-ceiling demote may flip to standard-PASS depending on the governance walk), V3-EXQ-534 SD-016 cue terrain training (continues the 2026-04-25 Path 1 diversification-loss thread). Two new ERRORs – V3-EXQ-244a MECH-165 reverse replay diversity validation and V3-EXQ-530 ARC-016 precision-commit circuit – both SIGTERM cloud kills at 2026-05-06T09:28-09:29Z (coincident infrastructure event on ree-cloud-1 / ree-cloud-2 – already noted in the morning’s diagnose-errors staging report).
• Queue depth: 9 -> 0 (empty items: [] 2026-05-07T01:12Z). The reef-superseding wave + the diagnostic-ablation backlog all completed through the day, with V3-EXQ-535 SD-029 P1 target fix the last item to be claimed (DLAPTOP-4.local at 21:48Z); on completion the runner pushed an empty queue.
• Pending review: 4 -> 32 items (7 PASS / 14 FAIL / 11 runner-only ERROR/UNKNOWN/smoke). Pending_review was regenerated by governance.sh at 2026-05-06T21:57Z immediately after the active governance cycle (claim governance-2026-05-06T2156Z) opened. The new accumulation reflects the entire 2026-05-06 wave landing in pending_review faster than the prior governance cycle could clear, plus three carry-overs from the 2026-05-05 reef-superseding wave (V3-EXQ-454a / 452a / 525). The active governance cycle is mid-walk at this nightly read (~3h old, well within the 6-hour stale-after-hours window).
• Active governance cycle (claim governance-2026-05-06T2156Z, opened 2026-05-06T21:56Z): holds REE_assembly governance files (claims.yaml / claims.json / review_tracker.json / pending_review.md / promotion_demotion_recommendations.md / experiment_proposals.v1.json / substrate_queue.json) plus WORKSPACE_STATE.md and TASK_CLAIMS.json. Final decisions land downstream of this nightly snapshot. The two PASSes flagged above (EXQ-528 SD-029 comparator-trained, EXQ-533 MECH-102 harm-stream ablation) plus EXQ-534 SD-016 cue terrain training are the obvious promotion-candidates the walk lands first; the FAILs distribute across SD-005 (532) / SD-015 (531) / SD-016 (418k) / Q-034 (526) / MECH-098 (529) / MECH-112 (527) / MECH-256+SD-029 (535).
• Bottleneck note: the 32-pending-review accumulation is the immediate gate. Underneath that, the monostrategy / reef-recovery thread remains the dominant scientific bottleneck – whether reef substrate alone is sufficient to recover behavioural acceptance under the SD-029 / SD-032b / SD-032c / Q-034 / MECH-112 / ARC-016 / MECH-257 / SD-049 Phase 2 cluster, or whether downstream calibration is also required. Three FAILs in the 2026-05-06 wave (V3-EXQ-526 Q-034, V3-EXQ-527 MECH-112 twice, V3-EXQ-535 SD-029 P1) sharpen that question – the reef substrate is in place but several downstream behavioural acceptance criteria are still not landing.

Immediate Work Queue (This Cycle, 2026-05-07)

1. Active governance walk completion (32 items) – 7 PASS to verify-and-close (V3-EXQ-418f SD-016 attention uniformity probe carry-over, V3-EXQ-523a SD-029 reef comparator, V3-EXQ-525 SD-003 attribution anchor PASS variant, V3-EXQ-528 SD-029 comparator-trained, V3-EXQ-244a MECH-165 – governance must classify since the run errored, V3-EXQ-533 MECH-102, V3-EXQ-534 SD-016 cue terrain); 14 FAIL to walk (the reef-superseding wave + V3-EXQ-525 + V3-EXQ-445g + V3-EXQ-517b + V3-EXQ-527 + V3-EXQ-526 + V3-EXQ-418k + V3-EXQ-532 + V3-EXQ-529 + V3-EXQ-531 + V3-EXQ-535); 11 runner-only ERROR/UNKNOWN/smoke to triage (V3-EXQ-244a, V3-EXQ-418j, V3-EXQ-445g, V3-EXQ-514d, V3-EXQ-514e, V3-EXQ-517b, V3-EXQ-524, V3-EXQ-530, V3-EXQ-531, V3-EXQ-533, V3-EXQ-534).
2. EVB-PINNED-Q019 backlog item – the pinned status note still reads “No literature extracted yet” (registered 2026-02-27). Its reading list has been fully satisfied since Feb and the corpus is now bias-corrected (4 weakens added 2026-05-06). A future cleanup should either un-pin or rewrite the status_reason; not blocking.
3. SD-049 Phase 2 z_resource encoder – still flagged ready=True priority=2 in substrate_queue.json. V3-EXQ-514c / 514d / 514e completed; whether the encoder identity expansion + phased training protocol P0/P1/P2 + SD-032 consumer cascade migration to read obs_dict['per_axis_drive'] are next is a governance-walk decision.
4. MECH-X commit_boundary + MECH-Y attribution_rigidity_setpoint registration – still pending user signoff on the 2026-05-03 commit-boundary-belief-lock lit-pull verdict (REGISTER WITH MODIFICATION).
5. V3-EXQ-490d Q-040b staleness-into-gate factorial – still drafted-but-unqueued pending design refinement.
6. V3-EXQ-495 V3-full-completion-gate re-run plan – still deferred until the Q-040b cluster-successor lands.
7. MECH-095 / MECH-102 substrate-enrichment review – SD-047 unblocks a successor experiment that exploits the multi-source dynamics substrate; the V3-EXQ-533 MECH-102 PASS may flip the substrate-ceiling classification on MECH-102. Governance walk decides.
8. EWIN-PC + ree-cloud-1/2 SIGTERM coincidence (2026-05-06T09:28-09:29Z) – two simultaneous cloud-worker kills (V3-EXQ-244a + V3-EXQ-530) plus one slightly earlier on V3-EXQ-495 are now a recurring pattern; supervisor / restart strategy on the cloud workers warrants follow-up.

---

Status Snapshot (2026-05-06T11:31Z – afternoon update, scheduled ree-lit-pull-pm: Q-019 disconfirming-balance lit pull + 627 runner completions + 18-item pending_review)

• Q-019 disconfirming-balance lit pull (2026-05-06T11:21Z, scheduled ree-lit-pull-pm). The pinned EVB-PINNED-Q019 backlog item was registered 2026-02-27 with a 6-paper reading list now satisfied; the genuine remaining gap was a confirmation-bias signature in the 16-entry corpus (11 supports / 5 mixed / 0 weakens). Pulled 4 disconfirming entries:
  • Redgrave / Prescott / Gurney 1999 (Neuroscience, 10.1016/s0306-4522(98)00319-4) – canonical unified-selection thesis (BG = single centralised selection device). conf=0.78.
  • Mink 1996 (Prog Neurobiol, 10.1016/s0301-0082(96)00042-1) – focused selection / surround-inhibit + focal-disinhibit on shared GPi/SNr output. conf=0.74.
  • Gurney / Prescott / Redgrave 2001 Part I (Biol Cybern, 10.1007/PL00007984) – GPR computational model: selection + control decomposition explicitly contra direct/indirect partitioning. conf=0.70.
  • Wouterlood et al. 2018 (J Neurosci Res, 10.1002/jnr.24242) – empirical cross-loop dopaminergic linkage from medial ventral striatum (limbic) to nigrostriatal cells projecting to dorsolateral striatum (sensorimotor); rules out strict three-loop segregation in the strongest form. conf=0.72.
• Effect on Q-019 evidence: entries 16 -> 20; direction_counts 11s/5m/0w -> 11s/5m/4w; lit_conf 0.884 -> 0.776; quadrant unchanged (plausible_unproven). Total lit entries 1179 -> 1183. The corpus is now properly balanced for governance to weigh option (A) single-gate against option (B) three-loop without the prior selection bias. REE_assembly pushed (75de72dc3).
• Substrate / SD status: no change since the morning 2026-05-06T01:35Z snapshot. Lit pull does not touch ree-v3 SD implementation, queue state, or runner_status.
• Runner activity since the morning 2026-05-06 snapshot: 608 -> 627 completions (+19 runs since the morning 04:28Z snapshot; cumulative breakdown 120 PASS / 255 FAIL / 72 ERROR / 180 UNKNOWN). Queue depth: 0 pending (everything has been claimed / completed); pending_review now lists 18 items (3 PASS / 10 FAIL / 5 runner-only ERROR/UNKNOWN/smoke). The pending_review backlog grew because the morning’s queued backlog (EXQ-528/531/533/534) plus the in-flight reef-superseding wave (514c/d/e, 517b, 523a, 526, 527) have been completing through the afternoon faster than they have been reviewed.

Immediate Work Queue (This Cycle, 2026-05-06 afternoon)

1. Pending_review walk (18 items) – 3 PASS to verify-and-close (V3-EXQ-418f SD-016 attention uniformity probe, V3-EXQ-523a SD-029 reef comparator, V3-EXQ-525 SD-003 attribution anchor); 10 FAIL to walk (the reef-superseding wave + V3-EXQ-525 + V3-EXQ-445g + V3-EXQ-517b + V3-EXQ-527 + V3-EXQ-526 + V3-EXQ-418k); 5 runner-only ERROR/UNKNOWN to triage via /diagnose-errors (V3-EXQ-445g, V3-EXQ-514d, V3-EXQ-517b, V3-EXQ-514e, V3-EXQ-418j).
2. Q-019 governance position – with 4 weakens entries now in the corpus, the question genuinely shifts from “three-gate vs single-gate” to “strict-segregation vs unified-with-topographic-inputs vs softer-multi-loop-with-cross-talk”. Strict segregation is the version Wouterlood 2018 directly disproves; the remaining viable hypotheses have smaller arbitration distance from each other than the existing corpus suggested. Q-019 was not promoted/demoted by this lit pull – the substrate-level question remains open and the next move is experimental, not bibliographic.
3. EVB-PINNED-Q019 backlog item – the pinned status note still reads “No literature extracted yet” (registered 2026-02-27). Its reading list has been fully satisfied since Feb and the corpus is now bias-corrected. A future cleanup should either un-pin or rewrite the status_reason; not blocking.

Status Snapshot (2026-05-06 – nightly docs sync, post-2026-05-04/05/06 reef-enrichment supersession wave: SD-019a / SD-051 / SD-052 / SD-050 reef substrate landings + reef-superseding monostrategy-blocked predecessors queued + INV-054 lit supplement + indexer backlog literature epoch-filter parity fix)

• SDs / MECHs moved to Implemented since the 2026-05-04 nightly snapshot:
  • SD-019a (harm_stream.immediate_affective_valence) IMPLEMENTED 2026-05-04 – third-tier z_harm_un EMA between fast z_harm_s and slow z_harm_a (~5-step rise at alpha=0.2); AIC + E3 short-horizon urgency redirect when use_harm_un; SD-021 descending modulation deliberately does NOT attenuate z_harm_un so controllability parity matches Loffler 2018; non-trainable EMA buffer; bit- identical OFF (184/184 contracts PASS with flag OFF). V3-EXQ-518 dry-run PASS, queued (4-arm, 9 acceptance criteria UC0a-b/UC1a-d/UC2a-b/UC3).
  • SD-051 / MECH-304 (safety_prediction.cue_specific_conditioned_inhibition_substrate) IMPLEMENTED 2026-05-04 – ConditionedSafetyStore non-trainable EMA prototype of z_world at MECH-302 event ticks + per-step decay forgetting; cosine similarity -> sigmoid -> commitment-release gate when beta elevated. New ree_core/safety/ package. V4-deferred items: approach attractor toward safety-signaling cues
    • contrastive cue-specific learning (require V4 multi-step planning + trainable encoder head). V3-EXQ-519 substrate-readiness queued; 6 of 9 sub-tests PASS, C6 surfaced upstream MECH-302 event-source dependency now under retest in V3-EXQ-517a/b.
  • SD-052 / MECH-303 (safety_prediction.contextual_passive_substrate) IMPLEMENTED 2026-05-04 – slow vmPFC/hippocampal-analog: ResidueField extended with safety_terrain_rbf_field + accumulate_safety + evaluate_safety; per-step accumulation when z_harm_a.norm() < harm_threshold AND not hypothesis_tag; commitment release when mean evaluate_safety >= release_threshold; same RBF pattern as ARC-030/MECH-117 benefit_terrain but separate field. V3-EXQ-520 4-arm substrate-readiness diagnostic dry-run PASS 10/10.
  • SD-050 reef enrichment substrate IMPLEMENTED 2026-05-04 in CausalGridWorldV2 – corner-adjacent Manhattan-radius reef safe zones (hazards excluded; 5x5 reef_field_view scent gradient appended to world_state, world_obs_dim 250->275) + food-attracted hazard drift bias (probability hazard_food_attraction). Two behavioral attractors – “flee to reef” vs “forage” – to break the single fixed route. V3-EXQ-521 substrate-readiness PASS 7/7; V3-EXQ-522 PASS zone_transitions=48.9 between attractors. NB: SD-050 ID is currently shared with the suffering-derivative comparator (also 2026-05-04); reconcile in a future cleanup.
• Strategic decision (2026-05-05): reef-enrichment supersession wave. The monostrategy-audit-2026-05-05T0712Z full-scan over failed experiments identified monostrategy as the dominant blocker across SD-029 / SD-032b / SD-032c / SD-049-Phase-2 / Q-034 / MECH-112 / MECH-257 / ARC-016 / SD-016. Rather than env-tune per EXQ, the lever is to swap in the SD-050 reef substrate as the env-entropy precondition under all affected predecessors. Reef-superseding versions queued: V3-EXQ-433e (SD-029 reef comparator), 445e/f/g (SD-032b dACC reef), 325f (SD-032c AIC descending reef), 452a (MECH-257 dual-function E2 reef), 454a (ARC-016 adaptive commitment reef), 514c (SD-049 Phase 2 reef behavioural validation), 526 (Q-034 reef threshold sweep), 527 (MECH-112 goal-directed reef + identity encoder). Three already FAILed and are on pending_review (454a, 452a, 525 SD-003 attribution anchor); whether reef alone is sufficient is the live question for the next governance walk.
• Indexer fix landed 2026-05-05: backlog literature-evidence epoch filter brought into parity with the matrix builder. One-line change at evidence/experiments/scripts/build_experiment_indexes.py:3002 mirroring the matrix’s stated Literature entries are not epoch-filtered policy. Symptom: morning lit-pulls kept flagging MECH-057, MECH-062, Q-019 as missing_literature_evidence despite well-populated targeted_review directories (lit_conf 0.78-0.89). Fix: if entry.get("source_type") == "literature" or is_applicable(entry):. Verification: governance.sh ran clean. MECH-057 lit_count 0->7, lit_conf 0.0->0.827, evidence_needed lit->experimental. MECH-062 dropped from backlog (confirmed_established). Q-019 pinned entry preserved. missing_literature_evidence reasons across backlog 3 -> 1 (only EVB-0131 onboarding phantom remains). Backlog item count 196 -> 205 because more claims now correctly register entries and trigger real signals (low_exp_conf etc) where previously they were silently lit-blank.
• Manifest correction wave 2026-05-05T22:15Z: EXQ-445g + EXQ-523a queued (dacc_bias_max_abs=2.0 fix supersedes 445f, evidence_direction_per_claim reset). EXQ-445f manifest evidence_direction_per_claim[“SD-032b”] corrected from “mixed” to “does_not_support” (entropy=0.0 all seeds, C2 wins=0). EXQ-523 flat JSON + per-run manifest evidence_direction corrected from “supports” to “non_contributory” with per-claim overrides + note + superseded_by added. Both run IDs marked reviewed in review_tracker.json. Governance pipeline ran clean (0 EXQ-523 warnings).
• Lit-pull supplement 2026-05-05: INV-054 depressive maintenance loop +3 entries (Jacobson 1996 BA RCT conf=0.78; van de Leemput 2014 PNAS bistable mood attractors conf=0.84; Tang 1999 sudden gains conf=0.76); INV-054 lit_conf 0.762 -> 0.858. LIT-0086 (ARC-029) and LIT-0088 (MECH-072) already well-covered (6 entries each), no new work needed.
• Experiment activity since the 2026-05-04 nightly snapshot:
  • +23 runner completions (585 -> 608; 113 -> 117 PASS / 248 -> 253 FAIL / 67 -> 70 ERROR / 157 -> 168 UNKNOWN). +4 PASS spans V3-EXQ-485a SD-033b multi-mode landing PASS, V3-EXQ-503 carry-over, V3-EXQ-521 reef substrate readiness PASS, V3-EXQ-522 reef behavioral diversity PASS.
  • Pending review queue regenerated 2026-05-05T22:12Z: 4 items. 3 FAIL (V3-EXQ-454a ARC-016 reef, V3-EXQ-452a ARC-033/MECH-257 reef, V3-EXQ-525 SD-003 attribution anchor) + 1 ERROR (V3-EXQ-418j SD-016 reef env-entropy fix; runner-only). The 2026-05-04T22:01Z governance walk had cleared pending_review after walking ARC-026 / MECH-093 promotions and the V3-EXQ-485a SD-033b multi-mode PASS; the 4-item residue is the post-walk accumulation from the reef-superseding wave + the SD-003 attribution anchor.

Immediate Work Queue (This Cycle, 2026-05-06)

1. Reef-superseding pending_review walk – 3 FAILs on the queue: V3-EXQ-454a (ARC-016 adaptive commitment under reef – whether reef restores the env-entropy needed for adaptive commitment to behave dissociably from fixed-threshold), V3-EXQ-452a (ARC-033 / MECH-257 / SD-013 dual-function E2 under reef – whether the comparator-vs-evaluator gating dissociation is recoverable under reef behavioural diversity), V3-EXQ-525 (SD-003 attribution anchor on post-SD-011 substrate – corrected output schema superseding EXQ-205). The governance question for each is: “Is reef alone the missing precondition, or does the architecture also need calibration?” The next session walks these decisions.
2. V3-EXQ-418j ERROR – SD-016 reef env-entropy fix runner-only ERROR; diagnose-errors should triage whether this is a reef substrate integration issue or a script-level bug.
3. Active in-flight queue (9 items) – V3-EXQ-524 (reef fishtank showcase episode log) + V3-EXQ-514c claimed by ree-cloud-2 (SD-049 Phase 2 reef behavioural validation) + V3-EXQ-514d claimed by DLAPTOP-4.local (BG gating diagnostic) + V3-EXQ-514e (BG gating seaweed diagnostic) + V3-EXQ-517b (MECH-302 relief-completion longer episodes, supersedes 517a) + V3-EXQ-523a (SD-029 reef comparator with adaptive graduation gate, supersedes 523) + V3-EXQ-526 (Q-034 reef threshold sweep, supersedes 451) + V3-EXQ-527 (MECH-112 goal-directed reef + identity encoder).
4. SD-049 Phase 2 z_resource encoder upgrade – still flagged ready=True priority=2 in substrate_queue.json. V3-EXQ-514c is the in-flight behavioural validation; once it lands, the encoder identity expansion + phased training protocol P0/P1/P2 + SD-032 consumer cascade migration to read obs_dict['per_axis_drive'] are next.
5. MECH-X commit_boundary + MECH-Y attribution_rigidity_setpoint registration – still pending user signoff on the 2026-05-03 commit-boundary-belief-lock lit-pull verdict (REGISTER WITH MODIFICATION).
6. V3-EXQ-490d Q-040b staleness-into-gate factorial – still drafted-but- unqueued pending design refinement in light of V3-EXQ-490c + V3-EXQ-490e FAILs.
7. V3-EXQ-495 V3-full-completion-gate re-run plan – still deferred until the Q-040b cluster-successor lands.
8. MECH-095 / MECH-102 substrate-enrichment review – SD-047 unblocks a successor experiment that exploits the multi-source dynamics substrate; design / proposal not yet drafted.
9. SD-051 / SD-052 substrate-readiness completion – V3-EXQ-519 6-of-9 PASS upstream-blocked by MECH-302 event source (under retest in V3-EXQ-517a/b); V3-EXQ-520 SD-052 dry-run PASS, awaiting full run.

---

Status Snapshot (2026-05-04 – nightly docs sync, post-2026-05-03 substrate-enrichment wave: SD-047 / SD-048 / SD-049 Phase 1 / SD-050 (MECH-302) substrate landings + ARC-026 / MECH-093 promotions + relief-completion + commit-boundary-belief-lock lit-pulls)

• SDs / MECHs moved to Implemented since the 2026-05-03 nightly snapshot:
  • SD-047 (environment.multi_source_dynamics) IMPLEMENTED 2026-05-03 in CausalGridWorld via flat kwargs (3 concurrent stochastic sources at distinct scales: AR(1) weather field + Poisson transient hazards + mobile drift sources; 4-arm noise-sweep lever; substrate-ceiling unblock for MECH-095 TPJ agency-detection; bit-identical OFF + per-source ablation; activation smoke ARM_2 calibration ratio 1.95:1 within 1:1-2:1 SD-doc target band). V3-EXQ-509 PASS 7/7 within minutes of queueing on Mac. SD-047 promoted candidate -> provisional, v3_pending removed.
  • SD-048 (body.interoceptive_noise_dynamics) IMPLEMENTED 2026-05-03 in CausalGridWorldV2 (3 concurrent agent-independent body-state noise sources on harm_obs_a readout: autonomic Gaussian + sensitisation Poisson + fatigue AR(1); Level 2 counterpart to SD-047 at body-state layer; substrate-ceiling unblock for ARC-058 / ARC-033 arbitration). V3-EXQ-511 6/7 with C1b sub-threshold at scale=0.25 – the 2026-05-03 governance walk accepted this as the SD-doc’s calibration-off-but- architecture-holds row (under-calibration on ARM_1, not architectural failure); evidence_direction reclassified weakens -> non_contributory with manifest note. SD-048 stays candidate / v3_pending pending V3-EXQ-512 ARC-058 comparator-gap behavioural successor.
  • SD-049 Phase 1 (environment.multi_resource_heterogeneity) IMPLEMENTED 2026-05-03 in CausalGridWorld (3 qualitatively distinct resource types incl non-homeostatic novelty channel + per-axis homeostatic drive vector parallel to legacy agent_energy via configurable max/mean/sum combiner + curriculum-introduction hook keyed on cross-episode _global_step; world_obs_dim 250 -> 325 default 3-type; benefit profiles sigmoidal_saturating / sharp_saturation / novelty_decay; substrate-roadmap H-priority #2; lit-anchored sd_049 lit_conf=0.898 across 5 PubMed entries). V3-EXQ-513 PASS 13/13 incl curriculum gates CC1/CC2 within minutes. Phase 2 z_resource encoder identity expansion + SD-032 consumer cascade
    • V3-EXQ-514 behavioural validation tracked in evidence/planning/substrate_queue.json as SD-049-PHASE-2 ready=True priority=2; v3_pending stays true on SD-049 until Phase 2 lands.
  • SD-050 / MECH-302 (relief.completion_event_reuses_goal_achievement_pipeline) IMPLEMENTED 2026-05-04 – SufferingDerivativeComparator on z_harm_a stream; reuses MECH-057a beta-gate release + MECH-094 categorical VALENCE_LIKING tag write; non-trainable; sense() ticks comparator, select_action() consumes event flag; bit-identical OFF; MECH-094 simulation_mode gate. V3-EXQ-515 comparator logic PASS 7/7 on Mac. V3-EXQ-516 agent-loop integration diagnostic queued (4-arm: ARM_0 OFF backward-compat / ARM_1 event fires / ARM_2 valence write / ARM_3 flat signal no false fires; ~1 min). MECH-302 / MECH-303 registered candidate / v3_pending earlier the same day from the relief-completion pre-registration lit-pull verdict.
• Promotions (2026-05-03T02:38Z governance walk):
  • ARC-026 (capacity_scaling.world_dim_cohort) candidate -> provisional. Two PASS pillars on different methodologies: EXQ-232 (full goal-conditioned agent) + V3-EXQ-507 (capacity-sweep weak-reading replication via random-action policy). conflict_ratio=0.286 below provisional gate.
  • MECH-093 (control_plane.zbeta_rate_modulation_dissociable) candidate -> provisional. Three PASS pillars: EXQ-097b step-level + EXQ-396b cross-experiment + V3-EXQ-505 substrate-level z_beta x salience factorial dissociation directly driving MultiRateClock. Resolves the EXQ-097 encoding-artifact conflict.
• Lit-pull wave 2026-05-03 (3 sessions, 23 PubMed-sourced entries):
  • Relief-completion mechanism (8 entries): Tanimoto 2004 + Andreatta 2012 LOAD-BEARING + Navratilova 2012 + Bromberg-Martin 2010 + Brischoux 2009 + Ramirez 2015 (M1 reward-circuit substrate for relief event); Kreutzmann 2020 + Meyer 2019 (M2/parallel substrate for safety-cue prediction). Verdict: hybrid-leaning-Model-1 conf 0.80. Outcome: MECH-302 (relief-completion event reuses goal-achievement pipeline) + MECH-303 (safety-cue parallel predictive substrate) both registered candidate / v3_pending in claims.yaml; 8 entries cross-tagged. MECH-302 lit_conf=0.904; MECH-303 lit_conf=0.780.
  • SD-049 multi-resource heterogeneity pre-implementation lit-pull (5 entries): Berridge 2018 + Smith & Berridge 2007 + Kidd & Hayden 2015 LOAD-BEARING (non-homeostatic novelty channel) + Shutts/Spelke 2009 + Matthews/Tye 2016 falsifier-adjacent (V3/V4 division of labor for social wanting). SD-049 lit_conf 0.0 -> 0.898.
  • Commit-boundary belief-lock pre-registration lit-pull (7 entries): Izuma 2010 PNAS conf 0.86 + Colosio 2017 J Neurosci conf 0.78 (cognitive dissonance + post-decisional consolidation); Saravanan 2010 BJP conf 0.68 (insight in psychosis – evidence indirect because clinical literature does not separate enacted from non-enacted delusional content); Gudjonsson 2016 Cortex conf 0.80 (forensic / pressured-internalized false confession with diary case); Sterzer 2018 Biol Psychiatry conf 0.84 (predictive-coding orthogonality cross-check explicitly flagging emergence-vs-persistence asymmetry as unresolved); Voigt 2018 J Neurosci conf 0.88 LOAD-BEARING + Tandetnik 2021 Cortex conf 0.82 (neural substrate of revising acted-upon vs non-acted-upon beliefs; cleanly DISSOCIATE two substrates – commit-record strength via hippocampus AND executive-conflict-resolution gain via frontal cortex). Verdict: REGISTER WITH MODIFICATION as two-mechanism cluster MECH-X commit_boundary_belief_lock + MECH-Y attribution_rigidity_setpoint (two-parameter substrate where pathological lock requires both substrates engaged at high gain, while pathological permeability / confabulation territory requires commit-record weakness regardless of executive engagement). All 7 entries tagged placeholder MECH-CBBL-PROPOSED to avoid inflating registered-claim lit_conf. Awaiting user signoff before MECH registration.
• Governance walks 2026-05-03 (two sessions):
  • 2026-05-03T02:38Z walk (post-V3-EXQ-504..508 cohort): 5 indexed pending walked; 2 PASS / 3 FAIL on raw scoring; 4 reclassifications applied; 2 promotions (ARC-026, MECH-093 – see above); SD-023 hold continued; MECH-057b / MECH-263 / Q-037 V3-pending hold continued; pending_user 5 -> 0; pending_review 6 -> 0; 8 manifest files edited, claims.yaml status+EQN updates for ARC-026 and MECH-093, 6 decision_log entries appended.
  • 2026-05-03T23:56Z walk (post-substrate-enrichment wave): 3 substrate- readiness diagnostics walked + 1 deferred runner-only ERROR. EXQ-509 SD-047 PASS 7/7 (clean close); EXQ-511 SD-048 FAIL 6/7 accepted as non_contributory per SD-doc interpretation grid; EXQ-513 SD-049 Phase 1 PASS 13/13; V3-EXQ-495 MECH-163 deferred (SIGTERM infrastructure kill on ree-cloud-1 at 4h of ~40h, exit code -15; defer until Q-040b cluster-successor lands). MECH-302 / MECH-303 / SD-049 pending_user holds applied (all hold_pending_v3_substrate, correctly held). 0 pending_user remaining. review_tracker updated, substrate_queue updated, claims.json rebuilt.
• Experiment activity since the 2026-05-03 nightly snapshot:
  • +11 runner completions (574 -> 585; 113 PASS unchanged / 247 -> 248 FAIL / 66 -> 67 ERROR / 148 -> 157 UNKNOWN). The +11 covers V3-EXQ-509/511/513/515 substrate-readiness PASSes (UNKNOWN runner result codes with PASS manifest verdicts), V3-EXQ-495 ERROR on ree-cloud-1, and indexer-surfaced carry-overs from the dedup sweep.
  • Pending review queue regenerated 2026-05-04T01:17Z: 2 items (both PASS). Both 2026-05-03 governance walks cleared their indexed pending sets at write time; the 2026-05-04 nightly indexer surfaced two PASS runs that completed after the 23:56Z walk: V3-EXQ-512 (SD-048 ARC-058 / ARC-033 comparator-gap behavioural successor; deferred at the 23:56Z walk but ran on Mac 2026-05-04T00:57Z and PASSed) and V3-EXQ-515 (MECH-302 suffering-derivative comparator substrate readiness; PASS indexer-surfaced 2026-05-04T01:17Z). Only V3-EXQ-516 (MECH-302 agent-loop integration diagnostic) remains queued; SD-049 Phase 2 follow-on remains the next substrate task with ready=True.

Immediate Work Queue (This Cycle, 2026-05-04)

1. V3-EXQ-516 result review – the queued MECH-302 agent-loop integration diagnostic (~1 min) is the only currently-queued experiment. Once it runs, governance walks the 4-arm acceptance (ARM_0 OFF backward-compat / ARM_1 event fires / ARM_2 valence write / ARM_3 flat signal no false fires) and decides whether MECH-302 is ready to flip from candidate / v3_pending to provisional, or whether a behavioural successor is needed.
2. SD-049 Phase 2 z_resource encoder upgrade – the highest-priority substrate task in the queue (substrate_queue.json SD-049-PHASE-2, ready=True, priority=2). 5-step phase_2_scope: (a) z_resource encoder identity expansion (one-hot or learned embedding); (b) phased training protocol P0/P1/P2; (c) SD-032 consumer cascade migration to read obs_dict['per_axis_drive']; (d) SD-012 emergent-invariant pending_substrate_reconfirmation flag; (e) V3-EXQ-514 4-arm behavioural validation pre-registered acceptance (goal_resource_r lift + identity-recovery probe + wanting != liking trajectory dissociation, with Woo/Spelke-style falsifier branch routing MECH-229 to substrate_conditional with V4-1 dependency on flat-failure).
3. MECH-X commit_boundary + MECH-Y attribution_rigidity_setpoint registration – pending user signoff on the 2026-05-03 commit-boundary-belief-lock lit-pull verdict (REGISTER WITH MODIFICATION as two-mechanism cluster). Once registered, the 7 placeholder-tagged entries cross-tag to the new claims and the per-mechanism lit_conf surfaces in the indexer.
4. V3-EXQ-490d Q-040b staleness-into-gate factorial – still drafted- but-unqueued pending design refinement in light of the V3-EXQ-490c + V3-EXQ-490e FAILs. The use_vs_gate_staleness_lookup OFF vs ON contrast at matched 0.4 thresholds remains the falsifiable test of the Q-040b strong reading via the 2026-04-29 MECH-284 wiring.
5. V3-EXQ-495 V3-full-completion-gate re-run plan – the 2026-05-03 ree-cloud-1 SIGTERM ERROR (4h of ~40h) was governance-deferred until the Q-040b cluster-successor lands. Re-run plan needs a more robust compute target (longer-running cloud capacity or ree-cloud-2 / EWIN-PC) given the ~40h horizon.
6. MECH-095 / MECH-102 substrate-enrichment review – now that SD-047 is implemented and V3-EXQ-509 PASSed, the MECH-095 substrate-ceiling unblock path is open: a successor experiment that exploits the multi-source dynamics substrate (rather than re-running V3-EXQ-506 on the legacy substrate) is the right next step. Design / proposal not yet drafted.
7. SD-048 V3-EXQ-512 ARC-058 comparator-gap behavioural successor – deferred from the 2026-05-03 governance walk pending the evidence_direction: non_contributory partial-PASS interpretation. The body-noise calibration ratio is now within the SD-doc target band at ARM_2 (2.39:1); the behavioural test of whether the comparator gap widens with the noise substrate ON is the architectural question.
8. V3 enrichment H-priority queue – substrate_roadmap.md still flags foreclosure primitives, multi-resource heterogeneity (Phase 2), and long-horizon dynamics as H-priority. Multi-resource Phase 1 landed 2026-05-03; Phase 2 is next-up. Foreclosure primitives + long-horizon dynamics remain at design-doc / SD-candidate stage.

---

Status Snapshot (2026-05-03 – nightly docs sync, post-2026-05-02 governance walk: SD-011 stable + SD-012 provisional + Phase-3-wave-2 epistemic_category schema + substrate_roadmap + v4_spec + Q-037/038/039 lit-pull)

• SDs / MECHs moved to Implemented since the 2026-05-02 nightly snapshot: none in the substrate sense – all substrate work landed in earlier waves. The 2026-05-02 work is governance / evidence / planning-artifact work layered on top of the existing run set.
• Promotions (2026-05-02T09:30Z queue walk):
  • SD-011 (harm_stream.dual_nociceptive_streams) provisional -> stable. exp_conf=0.871, lit_conf=0.871, 7 exp + 23 lit, conflict_ratio=0.148 (under the 0.20 stable gate). 25 supports / 2 weakens / 3 mixed across the largest single-claim corpus in the registry. EXQ-178b (harm_fwd_r2=0.742) and EXQ-323a body-damage substrate cleared the dissociation that had held it at provisional.
  • SD-012 (environment.homeostatic_drive) candidate -> provisional. exp_conf=0.714, lit_conf=0.874, 5 exp + 16 lit, conflict_ratio=0.20. 18 supports / 2 weakens / 1 mixed. Substrate has been implemented since 2026-04-02 (drive_weight=2.0 default in REEConfig.from_dims) – this promotion is the registry catching up to the substrate.
  • Bug fix included as part of this walk (commit 81de5101c, 2026-05-01): indexer recommendations function was counting scoring_excluded entries (diagnostic_probe / non_contributory / superseded) into the gate’s exp_conf computation. Phase 3 cutover surfaced this – spurious promotion recommendations for 5 claims with exp_conf=0 in the matrix were appearing pre-fix. After fix: pending_user 17 -> 15. After SD-011
    • SD-012 promotions: 15 -> 13.
  • Decision_log appended 2 entries (SD-011 stable, SD-012 provisional) with full rationale citing exp/lit/conflict numbers and the gate criteria each cleared. validate_claims –strict still PASSes 68/68 invariants.
• Phase 3 wave 2 governance schema landed 2026-05-02T11:30Z (epistemic_category formalisation):
  • New epistemic_category field on claims.yaml, OPTIONAL, with 7 canonical values: standard, substrate_coherence, answer_state, substrate_ceiling, substrate_conditional, derivational, out_of_domain.
  • _resolve_epistemic_category(claim_type, invariant_type, explicit) helper in build_experiment_indexes.py: explicit value wins; falls back to Phase 2 inference (architectural_commitment + universal-invariant -> substrate_coherence; open_question -> answer_state; else standard). Smoke-tested 10/10 cases.
  • _recommendation_for_claim dispatches via the resolver: only standard runs exp_conf gates; everything else suppresses promote/demote; narrow_open_question fires only for answer_state (so derivational / out_of_domain Q-claims stop getting “narrow this” recommendations).
  • _load_claim_registry parses optional epistemic_category field. validate_claims.py warn-only validates explicit values against the canonical set; invalid falls back to inference (does not crash).
  • 13 annotated claims backfilled with explicit categories: MECH-095, MECH-102 -> substrate_ceiling; Q-025, Q-026, Q-027 -> derivational; Q-028, Q-029 -> substrate_ceiling; Q-030 -> standard; Q-031, Q-032 -> out_of_domain; Q-037, Q-038, Q-039 -> substrate_conditional.
  • Result: pending decision queue dropped from 16 to 4 items (all 13 annotated claims correctly suppressed by their category; remaining items are pre-existing V3-pending holds + SD-023 conflict alert). validate_claims –strict still PASS 68/68.
  • CLAUDE.md “Claim-type evidence gating” section rewritten as “Epistemic categories (Phase 3 wave 2)” with the full mapping table and dispatch consequences.
• Substrate roadmap planning artifact landed 2026-05-02 (docs/architecture/substrate_roadmap.md):
  • Gathers V3 enrichment work into one planning document.
  • Documents 10 in-flight enrichments landed 2026-04-01..2026-05-02 (SD-022, SD-023, SD-029, SD-035, SD-036, SD-037, MECH-269 family, sleep aggregator, ghost-goal substrate, MECH-295 liking-bridge).
  • Enumerates 7 outstanding V3 enrichments:
    • 3 H-priority: foreclosure primitives, multi-resource heterogeneity, long-horizon dynamics.
    • 2 M-priority: multi-source environmental dynamics, differentiated coping channels V3-lite.
    • 2 L-priority.
  • Each substrate feature mapped to the claims it would unblock and the SD candidate that would register it.
• V4 spec planning artifact landed 2026-05-02 (docs/architecture/v4_spec.md):
  • Initial deliberate spec-first artifact for V4 substrate work.
  • 4 V4 primitive additions:
    • V4-1 multi-agent ecology
    • V4-2 self-model integration (DR-10..DR-14 from ree-v3 CLAUDE.md)
    • V4-3 long-horizon dynamics + persistent identity
    • V4-4 richer action repertoire
  • V4-bound claim cohort: ~12 claims explicitly waiting for V4 substrate.
  • Process gating: Phase A (this document, draft 2026-05-02). Phase B onwards gated on V3 full completion (MECH-163 PASS) + governance authorization.
  • Migration sketch: V4 substrate is additive; V3 continues for V3 claims; V4 evidence carries distinct architecture_epoch.
• Substrate-ceiling annotation walk 2026-05-02T10:35Z (13 claims annotated, no status changes): MECH-095 (TPJ agency-detection, 9 exp / 7 sup / 4 wk / 5 mix, conflict=0.727) and MECH-102 (violence-as-terminal, 24 exp / 10 sup / 11 wk / 12 mix, conflict=0.952) flagged as substrate_ceiling – both held at active. Diagnosis: CausalGridWorldV2 too coarse to deliver the distinctions these claims assert. 11 Q-claims (Q-025..032 / 037..039) walked and categorised per the mapping above. Pattern across the cohort: existing narrow_open_question recommendation collapses 4-5 distinct epistemic situations with different next-step implications – now machine-readable via the epistemic_category field.
• Weekend lit-pull 2026-05-02T10:00Z (Q-037 / Q-038 / Q-039, 9 PubMed-sourced entries):
  • Q-037 (psychosis substrate dissociability between MECH-094 / MECH-222 / MECH-065+223): Lavalle 2020 (dissociable source-monitoring OCD vs SCZ, supports MECH-094/MECH-222 dissociation), Corlett 2025 (20-yr aberrant salience review; argues MECH-065 may need to split into multiple sub-mechanisms), Oulton 2018 (PTSD memory amplification reality monitoring, supports MECH-094 tag-loss). lit_conf 0.0 -> 0.823.
  • Q-038 (D_V temporal-depth representational status): Pilkiw 2018 (distributed tonic+phasic temporal codes, supports option B), Guan 2024 (abstract interval-invariant subsecond structure via double training, supports option A), Gherman 2018 (VMPFC localised early confidence signal, supports option A by analogy). lit_conf 0.0 -> 0.795.
  • Q-039 (neuromodulators / control-plane vars regulating TCL integration window): Fan 2020 (ACh-driven L1 winner-take-all + temporal high-pass filter), Xiang 2023 (LC NA neurons phase-lock to infra-slow rhythms organising gamma), Kumagai 2023 (VNS double-dissociation: ACh selectively modulates gamma/beta, NA selectively modulates theta – cleanest dissociation evidence on the table). lit_conf 0.0 -> 0.84.
  • All three claims quadrant: plausible_unproven (high lit, no exp).
  • Indexer ran: literature_entries 1113 -> 1122.
• /diagnose-errors cleanup 2026-05-02T18:06Z: marked 3 obsolete unaddressed-error entries discussed in review_tracker.json (V3-EXQ-008 obsolete March SD-003 era; V3-ONBOARD-smoke-EWIN-PC contributor onboarding for inactive machine; V3-ONBOARD-smoke-ree-cloud-1 Hetzner CX22 onboarding for inactive cloud machine). Remaining 2 unaddressed errors confirmed intentional NotImplementedError sentinels (V3-EXQ-455a + V3-EXQ-449c) waiting on V3-EXQ-476 cascade gate + named consumer wiring – not bugs.
• Experiment activity since the 2026-05-02 nightly snapshot:
  • No new runner completions. runner_status.json totals unchanged at 574 (113 PASS / 247 FAIL / 66 ERROR / 148 UNKNOWN).
  • Pending review queue regenerated 2026-05-02T18:08Z: 1 item (V3-EXQ-490e FAIL on Q-040, carried over from 2026-04-30 governance walk; V3-EXQ-503 PASS reviewed 2026-05-01T20:40Z in the Phase 3 cutover session; 3 obsolete ERROR entries cleared 2026-05-02T18:06Z).

Immediate Work Queue (This Cycle, 2026-05-03)

1. V3-EXQ-490e FAIL discussion – still the last open pending_review item. Floor-relaxation arm has been ruled out as a Q-040b recovery path (combined with the V3-EXQ-490c FAIL on the matched-smoke-threshold factorial); the next-up falsification of the Q-040b strong reading is the staleness-into-gate test (V3-EXQ-490d toggle of use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds via the 2026-04-29T11:00Z MECH-284 wiring).
2. MECH-095 / MECH-102 demotion recommendations – now flagged substrate_ceiling (Phase 3 wave 2 schema); demotion-action suppressed by the new gating, but both still need a careful substrate-enrichment vs status-revision review at the next governance walk. The right response is substrate enrichment rather than more experiments on the existing substrate.
3. V3-EXQ-490d successor design + queueing – still the highest-priority substrate-validation run after the V3-EXQ-490e FAIL. Design refinement needed in light of the 490e FAIL before queueing.
4. V3-EXQ-495 V3-full-completion-gate queueing decision – still deferred. The MECH-163 dual-systems test depends on Q-040 / Q-040b resolution; queueing locked behind the cluster-successor outcome.
5. EXP-0174 env-complexity-gate scripting + queueing – the 2026-04-29T19:09Z proposal is unblocking for SD-016 retest path (env-entropy precondition) plus sleep / self-model aggregation retests (MECH-273 / MECH-275). Manual_proposals entry reserved; script not yet written.
6. OCD Layer 2 / 3 escalation (MECH-290 ablation; SD-046 multi-slot GoalState pull-forward) – still on the live escalation list after EXQ-498 disconfirmed Layer 1; design / queueing not yet started.
7. Aggregator-floor flag governance review – now a 6th-consecutive- cycle flag at the next governance walk; cap-aware aggregator review with recommendation either (a) accept the floor as architecturally reasonable for narrow-open-question Q-claims, or (b) tune the floor downward to expose per-paper confidence variance more faithfully.
8. V3 enrichment H-priority queue – substrate_roadmap.md flagged foreclosure primitives, multi-resource heterogeneity, and long-horizon dynamics as H-priority outstanding. None scheduled yet; design-doc / SD-candidate registration is the gating step.

---

Status Snapshot (2026-05-02 – nightly docs sync, post-Phase-2-cohort-closure + Option-E-Phase-3-cutover + duplicate-manifest-sweep)

• SDs / MECHs moved to Implemented since the 2026-04-30 nightly snapshot: none in the substrate sense – the substrate work is stable. The 2026-05-01 wave is governance / evidence / production-gate cutover work, plus the Phase 2 cohort closing experiment (SD-017 discriminative pair).
• Phase 2 cohort CLOSED on the experimental-evidence side (2026-05-01T20:15Z; reviewed 2026-05-01T20:40Z):
  • V3-EXQ-503 / EXP-0171a (SD-017 sleep-phase discriminative pair) PASS 3/3 seeds on Mac (~0.32s; runner result code UNKNOWN but manifest verdict PASS). Closes the SD-017 evidence gap left by V3-EXQ-500 – the substrate-readiness probe was diagnostic_probe and excluded from scoring, so SD-017 sat at exp_conf=0.000 / plausible_unproven despite lit_conf=0.901. Three pre-registered metrics M1 cumulative_sws_writes (ARM_A>=N_CYCLES, ARM_B==0); M2 ctxmem_state_change Frobenius norm (ARM_A>=0.10, ARM_B<=1e-6 – magnitude probe agnostic to slot-diversity-direction since SWS write can REDUCE diversity by clustering slots around real prototype geometry; the dry-run originally treated direction as the signal and FAILed C1 inverted, redesign was to switch to magnitude); M3 cumulative_rem_rollouts (ARM_A>=N_CYCLES, ARM_B==0). ARM_A: sws_writes=24, ctxmem_delta=4.50-5.03, rem_rollouts=18 across all seeds; ARM_B: zeros across the board.
  • Governance impact: SD-017 quadrant flipped plausible_unproven -> confirmed_established; exp_conf 0.000 -> 0.775. All four EXP-0170/171/171a/ 172/173 Phase 2 standard-gating cohort claims (MECH-094, SD-017, SD-035, MECH-062) are now confirmed_established under the new Phase 3 production gates.
• Lit/Exp Decoupling (Option E) Phase 3 cutover landed 2026-05-01: promotion / demotion gate logic now drives on experimental_confidence directly instead of the legacy overall_confidence blend. Specifics:
  • decision_criteria.v1.yaml: min_overall_confidence -> min_exp_conf, max_overall_confidence -> max_exp_conf. Legacy keys still accepted via the _t(d, new_key, legacy_key, default) helper for one-cycle backwards compat.
  • planning_criteria.v1.yaml: retired low_overall_confidence: 0.55; replaced with low_exp_conf: 0.55 and lit_only_above_cap: 0.50.
  • build_experiment_indexes.py:_decision_for_claim reads claim_meta["experimental_confidence"] directly. Recommendation rationale strings now report exp_conf=…, lit_conf=…, overall_confidence_legacy=….
  • Claim-type evidence gating brought INTO production (was Phase 1 shadow- only). substrate_coherence (ARC + universal invariant) and answer_state (open_question) skip exp_conf-based promote/demote – they fire conflict-resolution alerts and narrow_open_question only. standard gating (mechanism_hypothesis / design_decision / implementation / emergent or grey_zone invariant) fires the full set. Indexer _load_claim_registry now parses invariant_type from claims.yaml.
  • CLAUDE.md rewrote “Lit/Exp Decoupling Shadow” as “Lit/Exp Decoupling (Option E) – Phase 3 Cutover Done 2026-05-01” with the full three-phase history.
  • Diff against pre-cutover snapshot of promotion_demotion_recommendations.md: +2 actionable demotion recommendations surfaced (MECH-095, MECH-102 – both mechanism_hypothesis whose lit_conf was masking insufficient exp_conf under the legacy blend); 0 prior recommendations lost. All 4 Phase 2 standard-gating claims still confirmed_established under the new gates (MECH-094 exp=0.770; SD-017 exp=0.775; SD-035 exp=0.770; MECH-062 exp=0.770). Q-claim narrow_open_question recommendations and ARC conflict-resolution alerts all preserved.
  • Quadrant distribution after cutover: 194 plausible_unproven; 68 confirmed_established; 3 speculative; 1 novel_discovery.
• Duplicate-manifest sweep landed 2026-05-01 (31 phantoms / 25 clusters):
  • Phase 1 – 8 Tier-1 clusters (span < 2h between identical-signature emissions). Latest emission kept canonical, older copy marked evidence_direction=superseded with span-in-minutes and identical-signature sha1 in note. Claims that should see conflict_ratio movement on next index build: MECH-220, SD-012, MECH-112, ARC-032, MECH-116, MECH-090, ARC-016, SD-005, MECH-071, INV-034.
  • Phase 2 – 10 Tier-2 clusters (span 2-24h). 6 of 10 had no intervening commit on the experiment script and were auto-superseded; the other 4 had substantive commits between emissions and were flagged for manual review at evidence/experiments/dedup_review/phase2_manual_review.md (074f, 497, 496, 223). Claims that should see conflict_ratio movement: ARC-007, SD-004, MECH-033, MECH-072, SD-003, SD-007, SD-008, MECH-096, ARC-023.
  • Phase 3 – 10 Tier-3 clusters traced to runner regex bug active 2026-03-27..03-30, fixed by commit 071f1fc. Bug caused runner to mis-parse “Done. Outcome: PASS/FAIL” as UNKNOWN, leading to silent re-runs of completed experiments. Convention: kept OLDEST emission (legitimate original observation), superseded all later emissions (regex-bug-period replays). 12 phantom entries marked. Claims that should see conflict_ratio movement: ARC-032, MECH-116, MECH-112, MECH-117, Q-007, ARC-024, MECH-071, SD-003, MECH-111, Q-021, MECH-113, MECH-118, MECH-119, Q-022, ARC-029, ARC-030.
  • Per-run manifest mirroring – URGENT FIX: Phase 1/2/3 supersessions were originally applied to flat-JSON files at <exp_type>/<run_id>.json but the indexer reads <exp_type>/runs/<run_id>_v3/manifest.json. Mirrored the supersession decisions to per-run manifests so the indexer actually sees them. 31 per-run manifests across 25 clusters; 5 of 31 EXQ-232 ARC-026 manifests required re-application as their flat JSONs had been wiped by runner auto-sync between 2026-04-29 and 2026-05-01.
  • Phase 5 dedup guard – added _detect_and_mark_duplicate_emissions() to build_experiment_indexes.py as in-memory dedup guard with back-off when manual supersessions present; WARN logs emitted to stderr per duplicate. Smoke test passed (4s, 31 dups caught across 13 experiment_types, mostly old V2 epoch-excluded synthetic runs plus 4 new v3 clusters worth manual review: 074f, 133, 223, 484).
• Experiment activity since the 2026-04-30 nightly snapshot:
  • Runner-status totals refreshed: 574 completed (113 PASS / 247 FAIL / 66 ERROR / 148 UNKNOWN; +2 vs 2026-04-30 = V3-EXQ-490e FAIL + V3-EXQ-503 EXP-0171a SD-017 discriminative PASS). V3-EXQ-490e MECH-295 seeding-strengthening successor (Q-040b BASELINE-vs-RELAXED on activation- floor + drive_to_liking_gain knobs with MECH-295 bridge ON in both arms) FAIL on Mac 2026-05-01T03:19Z (~6h) – floor-relaxation alone does not recover approach_commit; combined with the V3-EXQ-490c FAIL the Q-040b strong reading remains weakened.
  • Pending review queue regenerated 2026-05-02T08:51Z: 1 item (V3-EXQ-490e FAIL on Q-040). The 2026-04-30T20:54Z governance walk reviewed the four Phase 2 cohort PASSes (V3-EXQ-499/500/501/502) plus the V3-EXQ-490c FAIL, dropping pending_review to 0+0 before the 490e FAIL and the V3-EXQ-503 PASS re-populated it; only V3-EXQ-490e remains open (V3-EXQ-503 PASS reviewed 2026-05-01T20:40Z in the Phase 3 cutover session).

Immediate Work Queue (This Cycle, 2026-05-02)

1. V3-EXQ-490e FAIL discussion – last open pending_review item. Floor-relaxation arm has been ruled out as a Q-040b recovery path (combined with the V3-EXQ-490c FAIL on the matched-smoke-threshold factorial); the next-up falsification of the Q-040b strong reading is the staleness-into-gate test (V3-EXQ-490d toggle of use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds via the 2026-04-29T11:00Z MECH-284 wiring).
2. V3-EXQ-490d successor design + queueing – still the highest-priority substrate-validation run on the substrate side after the V3-EXQ-490e FAIL. Design refinement needed in light of the 490e FAIL before queueing.
3. V3-EXQ-495 V3-full-completion-gate queueing decision – still deferred. The MECH-163 dual-systems test depends on Q-040 / Q-040b resolution; queueing locked behind the cluster-successor outcome.
4. EXP-0174 env-complexity-gate scripting + queueing – the 2026-04-29T19:09Z proposal is unblocking for SD-016 retest path (env-entropy precondition) plus sleep / self-model aggregation retests (MECH-273 / MECH-275). Manual_proposals entry reserved; script not yet written.
5. OCD Layer 2 / 3 escalation (MECH-290 ablation; SD-046 multi-slot GoalState pull-forward) – still on the live escalation list after EXQ-498 disconfirmed Layer 1; design / queueing not yet started.
6. MECH-095 / MECH-102 demotion recommendations – newly surfaced by the Phase 3 cutover (both mechanism_hypothesis whose lit_conf was masking insufficient exp_conf under the legacy blend); both need review at the next governance walk. No actionable promotions or demotions other than these two surfaced by the cutover.
7. Aggregator-floor flag governance review – 5th-consecutive-cycle flag at the next governance walk; cap-aware aggregator review with recommendation either (a) accept the floor as architecturally reasonable for narrow-open-question Q-claims, or (b) tune the floor downward to expose per-paper confidence variance more faithfully.

---

Status Snapshot (2026-04-30 – nightly docs sync, post-Phase-2-claim-type-gating cohort PASS wave)

• SDs / MECHs moved to Implemented since the 2026-04-29 PM snapshot: none in the substrate sense (the same-day MECH-269b + MECH-284 staleness-into-VsRolloutGate wiring landed at 06:03Z and was already captured in the 2026-04-29 PM snapshot below). The day’s primary signal is on the experimental-evidence side: the 2026-04-29 PM Phase 2 claim-type-gating reckoning surfaced four standard-gating MECH/SD claims that needed first-pass experimental signal, all four scripts were written + smoke-tested + queued + executed in one session, and all four PASSed in 3/3 seeds.
• Phase 2 claim-type evidence gating landed (2026-04-29T15:47Z): scripts/generate_option_e_shadow.py extended with _evidence_gating(meta) classifying every claim into standard | substrate_coherence | answer_state based on claim_type (and invariant_type for invariants). Discrepancy / impl_no_exp / low_exp / lit_only_above_cap flags now fire only for standard-gating claims; the suppressed buckets surface separately for transparency. Headline numbers: discrepancies 183 -> 135 (48 ARC + universal invariant + Q dropped out of actionable), impl_no_exp 15 -> 4 (the 4 genuinely-testable MECH/SD claims that needed experiments), low_exp 50 -> 32, lit_only_above_cap 142 -> 107. CLAUDE.md “Claim-type evidence gating” section documents the rules + restate-Q-as-MECH/SD path.
• Phase 2 cohort experiments (4 PASS in 3/3 seeds; pending review):
  • V3-EXQ-499 / EXP-0170 (MECH-094 hypothesis-tag write-gate discriminative pair) – substrate-level forced-injection design avoiding the V3-EXQ-140 non_contributory failure mode. ARM_A passes hypothesis_tag=True for sim events (gate active per MECH-094); ARM_B passes False (gate ablated, the tag-loss / write-channel disinhibition pathology). N_REAL=50 + N_SIM=50 deterministic events at disjoint z_world locations. Three pre-registered metrics with PASS criteria all met >= 2/3 seeds: ARM_A contam=0.000, confab=0.000, MI=0.693 (perfect log(2)); ARM_B contam=1.000, confab=0.640, MI=0.000. Governance impact: MECH-094 quadrant flipped plausible_unproven -> confirmed_established; experimental_confidence 0.000 -> 0.775; overall_confidence 0.866 -> 0.843 (slight dip from legacy blend now weighting the new exp signal). First standard-gating experimental evidence for MECH-094 – prior 9 entries excluded as diagnostic_probe / non_contributory.
  • V3-EXQ-500 / EXP-0171 (SD-017 sleep-phase substrate-readiness diagnostic) – nine prior FAIL/non_contributory entries on SD-017 led to a fresh substrate-readiness gate before any behavioural retest. PASS.
  • V3-EXQ-501 / EXP-0172 (SD-035 amygdala analog vs binary) – discriminator between SD-035’s analog amygdala substrate and a degenerate binary toggle. PASS.
  • V3-EXQ-502 / EXP-0173 (MECH-062 tri-loop gate coordination) – truly fresh-start MECH-062 evidence (zero priors); first tri-loop coordination test on the V3 substrate. PASS.
  • manual_proposals.v1.json: EXP-0170/171/172/173 status -> executed.
• Experiment activity since the 2026-04-29 PM snapshot:
  • Runner-status totals refreshed: 572 completed (113 PASS / 246 FAIL / 66 ERROR / 147 UNKNOWN; +4 vs 2026-04-29 PM = V3-EXQ-499/500/501/502 Phase 2 cohort).
  • Pending review queue regenerated 2026-04-29T19:32Z: 5 items (4 PASS V3-EXQ-499/500/501/502 + 1 FAIL V3-EXQ-490c carried over).
• EXP-0174 env-complexity-gate proposal added to manual_proposals.v1.json at 19:09Z (high-priority diagnostic targeted_probe for SD-016 / SD-023 / MECH-273 / MECH-275 testing whether enriched environments produce enough z_world information for SD-016 retrieval and sleep / self-model aggregation retests; experiment_proposals + index regenerated; no queue entry reserved or appended).

Immediate Work Queue (This Cycle, 2026-04-30)

1. Governance walk on the 5 pending review items – the 4 Phase 2 cohort PASSes (V3-EXQ-499/500/501/502) now carry first-pass standard- gating experimental evidence for MECH-094 / SD-017 / SD-035 / MECH-062; re-score against the lit/exp decoupled regime (overall_confidence_decoupled = exp_conf) and check whether any quadrant flip is promotion-relevant. The V3-EXQ-490c FAIL remains pending discussion with successor V3-EXQ-490d already scoped.
2. V3-EXQ-490d successor design + queueing – still the highest-priority substrate-validation run on the substrate side. Toggle use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds, 2 arms x 3 seeds, MECH-269b V_s gating ON + MECH-295 bridge ON in both arms. C4 severance acceptance criterion is the falsifiable test of the Q-040b strong reading after the 490c FAIL.
3. V3-EXQ-495 V3-full-completion-gate queueing decision – still deferred. The MECH-163 dual-systems test depends on Q-040 / Q-040b resolution; queueing locked behind the EXQ-490d outcome.
4. EXP-0174 env-complexity-gate scripting + queueing – the 2026-04-29T19:09Z proposal is unblocking for SD-016 retest path (env-entropy precondition) plus sleep / self-model aggregation retests (MECH-273 / MECH-275). Manual_proposals entry reserved; script not yet written.
5. OCD Layer 2 / 3 escalation (MECH-290 ablation; SD-046 multi-slot GoalState pull-forward) – still on the live escalation list after EXQ-498 disconfirmed Layer 1; design / queueing not yet started.
6. Aggregator-floor flag governance review – 5th-consecutive-cycle flag at the next governance walk; cap-aware aggregator review with recommendation either (a) accept the floor as architecturally reasonable for narrow-open-question Q-claims, or (b) tune the floor downward to expose per-paper confidence variance more faithfully.

---

Status Snapshot (2026-04-29 PM – post-V3-EXQ-490c FAIL + Q-032 lit-pull + MECH-269b staleness wiring)

• SDs / MECHs moved to Implemented since the 2026-04-29 nightly snapshot: none in the substrate sense. One non-trivial wiring extension landed at 06:03Z: MECH-269b + MECH-284 staleness-into-VsRolloutGate wiring (a Q-040b strong-reading enabler; effective_vs = raw_vs - staleness[s] under use_vs_gate_staleness_lookup, with HippocampalModule.compute_per_stream_staleness aggregating max-over-active-anchors-whose-stream_mixture-includes-stream). Bit-identical to legacy raw-V_s path when flag OFF; 8/8 contract tests PASS (191/191 full preflight + contracts). Successor V3-EXQ-490d can now drop the smoke-threshold override and exercise C4 severance (use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds) as the falsifiable test of the strong reading.
• Experiment activity since the nightly snapshot:
  • V3-EXQ-490c (MECH-269b V_s gating + MECH-295 liking-bridge factorial; Q-040b behavioural sufficiency arm) completed FAIL on Mac 2026-04-29T08:34Z (~2.6h). Preliminary reading: under matched smoke- threshold overrides (0.85/0.85/0.95), MECH-269b ON + MECH-295 ON jointly do NOT recover approach_commit. Q-040b strong reading is NOT supported in this configuration. Pending review; governance decision pending.
  • Successor V3-EXQ-490d scoped (not queued): drop smoke-threshold override; toggle use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds in a 2-arm 3-seed factorial; C4 severance becomes the falsifiable test of whether MECH-284 staleness shifts effective_vs enough to fire the gate at biologically realistic V_s readings.
  • Runner-status totals refreshed: 568 completed (111 PASS / 245 FAIL / 66 ERROR / 146 UNKNOWN; +1 vs nightly = V3-EXQ-490c).
  • Pending review queue regenerated 11:36Z: 1 item (V3-EXQ-490c runner-only FAIL pending discussion).
• Lit-pull progress 2026-04-29 (3 sessions, 15 entries):
  • AM (08:41Z) Q-029 + Q-030: 5 entries each. Q-029 loneliness-as- ethical-harm-derivable-from-INV-029 lit_conf=0.875 (Holt-Lunstad x2, Wang 2023 Nat Hum Behav, Zajner/Bzdok UK Biobank, Eisenberger AnnRevPsych). Q-030 z_resource/z_world separation/fusion permutations lit_conf=0.874 (Staresina PrC/PhC, Lee/Inah GIST, Howard/Kahnt OFC identity vs vmPFC value, Kim DG-disrupted spatial-object binding, Locatello Slot Attention).
  • PM (11:19Z) Q-032: 5 entries; PSG SWS/REM ratio as pharmacodynamic biomarker for sleep-medication dementia outcomes; lit_conf=0.839. Limb (a) PSG predicts dementia at the individual level: Himali 2023 JAMA Neurol (Framingham within-person SWS-decline x 17y dementia HR 1.27, 0.82); Winer/Mander/Walker 2020 Curr Biol (NREM SWA <1Hz forecasts AB accumulation with cross-metric specificity, 0.78); Suh 2019 J Alzheimers Dis (KLOSCAD short REM-per-cycle predicts MCI conversion at 4y, 0.62 exploratory). Limb (a)/(c) multi-metric architectural signature: D’Rozario 2020 Sleep Med Rev (10-study meta-analysis MCI fingerprint, 0.72). Limb (b)/(c) field treats PSG SWS as legitimate trial endpoint: Eyob 2024 J Alzheimers Dis (REST trazodone protocol, 0.45). Synthesis verdict: limb (a) well-supported; limb (c) PD-biomarker proposal architecturally defensible; limb (b) drug-class differential SWS preservation -> differential cognitive outcomes still empirically open pending REST results.
  • Aggregator floor flag continues (4th consecutive cycle): per-paper confidences 0.45-0.82 averaging ~0.68; claim-level lands 0.83-0.88. Floor effect now visible across Q-027/Q-028/Q-029/Q-030/Q-031/Q-032. Worth flagging at the next governance walk for cap-aware aggregator review.
• Indexer state: literature entries 1098 -> 1113 across the day’s three lit-pull sessions (+15); claim_evidence.v1.json + evidence_backlog.v1.json regenerated 11:29Z.

Immediate Work Queue (This Cycle, 2026-04-29 PM)

1. V3-EXQ-490d successor design + queueing – the highest-priority substrate-validation run. Toggle use_vs_gate_staleness_lookup OFF vs ON at matched 0.4 thresholds, 2 arms x 3 seeds, MECH-269b V_s gating ON + MECH-295 bridge ON in both arms (inherits 490c stack). C4 severance acceptance criterion becomes the falsifiable test of the Q-040b strong reading after the 490c FAIL.
2. V3-EXQ-490c review + governance decision – the run sits in the pending-review queue (1 item). Decision tree: classify as inconclusive pending V3-EXQ-490d, or accept FAIL as evidence against the strong reading and re-route Q-040b to ghost-goal / planning-arm hypotheses.
3. V3-EXQ-495 V3-full-completion-gate queueing decision – still deferred. The MECH-163 dual-systems test depends on Q-040 / Q-040b resolution; queueing locked behind the EXQ-490d outcome.
4. OCD Layer 2 / 3 escalation (MECH-290 ablation; SD-046 multi-slot GoalState pull-forward) – the EXQ-498 disconfirmation of OCD Layer 1 leaves Layers 2 and 3 as the live escalation paths; design / queueing not yet started.
5. SD-016 env-entropy precondition resolution – SD-016 remains parked pending env-entropy precondition; broader env-enrichment scoping work (CausalGridWorldV2 extension) not yet started.
6. Aggregator-floor flag governance review – 5th-consecutive-cycle flag at the next governance walk; cap-aware aggregator review with recommendation either (a) accept the floor as architecturally reasonable for narrow-open-question Q-claims, or (b) tune the floor downward to expose per-paper confidence variance more faithfully.

---

Status Snapshot (2026-04-29 – nightly docs sync, post-2026-04-28 governance cycle)

• SDs / MECHs moved to Implemented since the 2026-04-28 PM snapshot: none. No new substrate landings; the day’s activity was dominated by the 2026-04-28T23:04Z governance cycle and a six-experiment diagnostic wave (V3-EXQ-498 OCD Layer 1 / V3-EXQ-418f-h SD-016 / V3-EXQ-490b MECH-269b) that produced no PASSing substrate signals but did shrink the open hypothesis space substantially.
• Governance promotions / status changes 2026-04-28T23:04Z (10 user- approved decisions across 20 walked items):
  • 4 promotions candidate -> provisional: MECH-266 (asymmetric per-mode hysteresis), MECH-267 (mode-conditioned hippocampal proposals), MECH-268 (dACC conflict saturation), SD-034 (governance closure operator).
  • SD-033b v3_pending true -> false based on V3-EXQ-485 substrate-landing PASS UC1-UC5; status remains candidate (behavioural MECH-263 devaluation + task-role discrimination signatures deferred to environment-extension EXQs).
  • 2 holds preserved as hold_pending_v3_substrate: MECH-057b (hippocampal candidacy gate; functional signature deferred) and MECH-263 (OFC functional signatures; signatures (a)/(b) explicitly DEFERRED in V3-EXQ-485 manifest notes).
  • 6 narrow-open Q-claim evidence_quality_note refreshes capturing the 2026-04-28 lit-pull cluster + 5th-consecutive aggregator-floor caveat: Q-025 / Q-026 / Q-027 / Q-028 / Q-031 / Q-040.
• Manifest edits applied this cycle:
  • V3-EXQ-485 (4 manifests) MECH-263 supports -> non_contributory per claim_ids accuracy rule (experiment tests SD-033b substrate wiring only; MECH-263 functional signatures explicitly deferred in manifest notes).
  • Replica supersession: kept canonical EXQ-484 (20260427T054449Z), EXQ-485 (20260427T054454Z), EXQ-493 (20260427T080304Z); marked 4+3+2 deterministic re-runs evidence_direction=’superseded’ to avoid 5x/4x/3x over-weighting.
  • V3-EXQ-418g SD-016 does_not_support -> non_contributory (env-entropy precondition gap traced via EXQ-418f probe + EXQ-418h FAIL; substrate-side fixes work as designed but cue_context produces zero behavioural delta because z_world is near-constant batch-wise).
• Diagnostic wave 2026-04-28 (6 experiments, all FAIL/UNKNOWN):
  • V3-EXQ-498 OCD Layer 1 closure-threshold sweep – non_contributory. Layer 1 hypothesis (sweeping closure_rule_delta_threshold attenuates V3 baseline monostrategy) DISCONFIRMED. Escalates to Layer 2 (MECH-290 ablation diagnostic) or Layer 3 (SD-046 multi-slot GoalState pull-forward).
  • V3-EXQ-418f SD-016 attention-uniformity probe – diagnostic only. Localised the EXQ-418d/418e ln(16) uniform-rail bottleneck to query selectivity (not slot content): A2_div_only achieved slot_diversity=0.9999 but attn_entropy stayed at the uniform rail and contributed zero behavioural delta vs A0_off.
  • V3-EXQ-418g SD-016 selectivity-first 4-arm – non_contributory. Substrate-side fixes (learnable temperature + entropy regulariser) work as designed: C1+C2+C3 PASS (B1/B3 attn_entropy=0.000, B2/B3 slot_diversity=1.000). C4+C5 FAIL with action_class_entropy~1.1e-10 IDENTICALLY across ALL FOUR arms because cue_context produces zero behavioural delta when z_world is near-constant across the batch (cos~0.998 in 418f).
  • V3-EXQ-418h env-entropy precondition probe – FAIL. SD-023 landmarks-on alone does NOT supply enough cross-context z_world variance under the current env (H1 cos_cross<0.95 not robust across seeds). Routes to broader env-enrichment scoping.
  • SD-016 parked pending env-entropy precondition. substrate_queue status: parked_pending_env_entropy_precondition. validation_experiment rerouted to EXQ-418h family pending broader env scoping.
  • V3-EXQ-490b MECH-269b VsRolloutGate substrate-readiness probe – FAIL outcome under UNKNOWN result code, governance classified inconclusive. Q-040a effective PASS at threshold-overridden smoke (vs_gate_e1/e2_threshold=0.85, snapshot_refresh=0.95 so the gate fires under typical Phase 1 V_s dynamics). Q-040b stale-stream- discrimination FAIL points at MECH-295 drive->liking->approach bridge as the still-open dependency. The MECH-295 bridge landed 2026-04-26 (V3-EXQ-493 PASS), so the successor combines MECH-269b VsRolloutGate ON with MECH-295 bridge ON in a single factorial. Successor EXQ pending design.
• Lit-pulls landed 2026-04-28 (PM, 5 EVB-IDs from morning agenda): 16 new entries across MECH-025b (3) + MECH-172 (5, new dir) + Q-019 (2 top-up; saturated dir) + INV-043 (3) + MECH-187 (3). Indexer rebuilt to 1098 lit entries (+16). Post-rebuild: MECH-025b lit=0.82 (was 0.0); MECH-172 lit=0.894 (was 0.0); Q-019 lit=0.887 (~unchanged, saturated); INV-043 lit=0.841 (was 0.65); MECH-187 lit=0.863 / overall=0.829. Morning-agenda label drift flagged (4 of 5 mismatched between agenda’s claim labels and resolved EVB-IDs); aggregator floor effect flagged for 5th consecutive cycle (per-paper 0.55-0.86 averaging ~0.74 but claim-level 0.82-0.89).
• Pipeline state: validate_claims –strict OK 68 invariants; claims.json rebuilt 571; index 915 runs / 482 types; pending_review 0+0. Substrate queue 52 items (32 implemented + 2 implemented_but_failing_validation + 1 parked + 17 pending/blocked). 10 proposals flipped to executed across MECH-261/262/263/295/SD-033a/ SD-033b/SD-034/MECH-266/267/268/Q-040.
• Experiment count: 567 runner-side completions per runner_status.json 2026-04-28T21:12Z read (111 PASS / 244 FAIL / 66 ERROR / 146 UNKNOWN). +6 over the 2026-04-27 read covering the 2026-04-28 diagnostic wave above.
• Pending review: 0 items per pending_review.md regenerated 2026-04-28T23:11Z (was 15 at 2026-04-28T04:18Z). Cleared by the 2026-04-28T23:04Z governance cycle.
• Queue (experiment_queue.json 2026-04-29): 1 item, claimed by ree-cloud-2.
  • V3-EXQ-490b completed 2026-04-28T21:09Z UNKNOWN/FAIL; awaiting queue cleanup. Governance-classified inconclusive; combined EXQ-490b/MECH-295 successor pending design.
  • V3-EXQ-495 drafted but not yet queued. THE V3-full-completion gate run; queueing decision deferred until the EXQ-490b/MECH-295 successor lands.
• Current first-paper-gate bottleneck: the EXQ-490b/MECH-295 combined successor is the next-up substrate-validation run. V3-EXQ-495 (V3 full- completion gate / MECH-163 hippocampally-planned arm) remains the headline first-paper-gate run. The EXQ-483 wired-but-inert pattern remains the open behavioural-recovery thread for the SD-037 / MECH-269b / MECH-295 cluster. SD-016 parked pending env-entropy precondition. OCD Layer 1 disconfirmed (escalation to Layer 2 / Layer 3).

---

Status Snapshot (2026-04-28 PM — afternoon docs sync, lit-pull wave + queue refresh)

• SDs / MECHs moved to Implemented since the 2026-04-28 nightly snapshot: none. No new substrate landings between 01:21Z and 11:30Z; runner_status.json unchanged at 561 completions (109 PASS / 242 FAIL / 66 ERROR / 144 UNKNOWN per the 2026-04-27T08:04Z read).
• Experiment count (per runner_status.json 2026-04-27T08:04Z): 561 total (unchanged since morning). Pending review queue regenerated 2026-04-28T04:18:29Z carries 15 items – 12 PASS (V3-EXQ-484 / 485 / 493 across multiple machine/timestamp runs indexed after the 2026-04-27T14:55 /diagnose-errors run_id naming-bug fix) and 3 runner-only UNKNOWN entries for the same three queue IDs awaiting next governance walk. The queue grew from 6 at the 2026-04-27T14:47:47Z regen because each per-machine/per-timestamp PASS now indexes as a distinct run; underlying queue IDs unchanged.
• Lit-pull wave 2026-04-28 (5 pulls totalling 25 entries; medium-priority backlog largely closed):
  • Q-031 (anticholinergic burden / dementia / REM-mediation): 5 entries, new dir targeted_review_q_031/ (PM scheduled task, this snapshot). Gray 2015 ACT cohort + Coupland 2019 QResearch (cumulative-dose limb both supports); Pase 2017 Framingham (REM% –> incident dementia, REM-specific not NREM); Kim & Jeong 1999 (transdermal scopolamine suppresses phasic REM, n=8 acute substrate); Grace-Vanstone-Horner 2014 J Neurosci mixed – pontine SubC ACh blockade does NOT abolish REM; weakens strong pathway-1, suggests REM-theta integrity may be the right mediator rather than REM%. lit_confidence=0.87 (was not in claim_evidence map). Q-031c mediation-analysis sub-question has no published direct test – flagged.
  • Q-027 / Q-028 (irreversible harm / axiom conflict): 5 entries (AM scheduled task; Sunstein 2006, Tarsney-Thomas-MacAskill 2024 SEP, IEP Precautionary, McConnell 2022 SEP, Williams 1965).
  • theta-abstraction-scaling: 5 entries – new MECH claims registered: MECH-299 (theta_cycle_content_scales_with_substrate_vocabulary; V4 firm), MECH-300 (cognitive_map_traversal_at_active_abstraction_level; V4 firm), MECH-301 (waking_quiescent_replay_via_mech285_priority_sampling; V4 default with V3 PULL-FORWARD CONDITION).
  • action-policy-decomposition: 5 entries (Mussa-Ivaldi 2000 / Daw 2005 / Graybiel 2008 / Botvinick 2009 / Dolan & Dayan 2013); synthesis verdict flags habit / action-chunk substrate (level 2-3 of biological decomposition) as missing in V3 – highest-priority extension for OCD modelling and potential monostrategy contributor. Filed as candidate SD.
  • hpc-type-prototype: 5 entries (Quiroga 2005 / Schapiro 2017 / Schapiro 2016 / Constantinescu 2016 / Hennies 2017). Verdict: parsimonious REE extension is a NEW INPUT PROJECTION onto existing AnchorSet machinery (Constantinescu shared-cognitive-map reading) plus an explicit prototype-readout operator running BOTH waking and sleep, NOT a fully separate codebook substrate.
• Aggregator floor flag (recurring; third consecutive surfacing today): per-paper confidences for low/medium-anchored Q-claims average 0.62-0.74 but claim-level lit_confidence aggregates to 0.85-0.88 (Q-031 0.87; Q-027 0.66; Q-028 0.63 morning). Worth flagging next governance walk for a cap-aware aggregator review – the floor effect is causing claim-level numbers to read as more lit-supported than the per-paper scoring intends.
• Queue refresh (3 new diagnostics queued today, 2026-04-28):
  • V3-EXQ-498 OCD Layer 1 closure-threshold sweep (SD-034 parameter diagnostic; 4-arm × 3 seeds; psychiatric_failure_modes.md OCD section Layer 1 hypothesis test; PASS/FAIL routes to Layer 2 / Layer 3 escalation). claim_ids=[‘SD-034’]. ~60 min.
  • V3-EXQ-418f SD-016 attention-uniformity diagnostic probe (single-seed instrumentation of EXQ-418d/e ln(16) uniform-rail bottleneck). ~15 min.
  • V3-EXQ-418g SD-016 Path 4 query-selectivity-first 4-arm with new learnable temperature + attention-entropy loss substrate hooks landed 2026-04-28; B0_off / B1_sel_only / B2_div_only / B3_sel_plus_div × 3 seeds; tests Path 4 hypothesis that query selectivity (not slot content) is the bottleneck. ~90 min.
• Current bottleneck unchanged: V3-EXQ-495 (V3 full-completion gate / MECH-163 hippocampally-planned arm) is the headline run; V3-EXQ-490b is the smaller upstream factorial for Q-040a; V3-EXQ-498 is now the proximate short-runtime diagnostic that can land before the 25h V3-EXQ-495 budget decision.

---

Status Snapshot (2026-04-28 — nightly docs sync, post-2026-04-27 substrate wave)

• SDs / MECHs moved to Implemented since the 2026-04-27 morning snapshot:
  • SD-039 module-level write-site population layer (hippocampal.anchor_goal_payload_population) – the deferred follow-on to the SD-039 substrate. REEAgent.sense() builds AnchorGoalPayload once per tick from GoalState (z_goal_snapshot), ResidueField VALENCE_WANTING (wanting_strength), BLA arousal_tag, mean(per_stream_vs) (last_vs), max staleness_accumulator (staleness_at_write), and threads it through both HippocampalModule.tick_anchor_set (boundary-event write/remap path) and apply_invalidation_broadcasts_to_regions (MECH-287 broadcast invalidation refresh). MECH-094 simulation-mode gate at build_goal_payload returns None on replay/DMN paths. REEConfig.from_dims(use_sd039_anchor_payload=...) propagates to AnchorSetConfig.use_sd039_anchor_payload. V3-EXQ-494 6/6 PASS (UC1-UC6: module importable; master OFF no-op; population fires 7/7 anchors with max goal_match 0.9999; dual-trace preservation 6 inactive
    • 1 active; falsifiable signature Phase A mean=0.0 vs Phase B mean=0.998 with 3/3 above 0.3; MECH-094 simulation gate replay-path zero anchors).
  • MECH-292 ranked ghost-goal bank (hippocampal.unresolved_goal_ghost_bank) – pure-arithmetic derived view over the SD-039 dual-trace anchor pool. New module ree_core/hippocampal/ghost_goal_bank.py (GhostGoalBank, GhostGoalBankConfig, GhostGoalBankEntry). Ranking formula ghost_priority = w_w*wanting + w_m*goal_match + w_s*staleness + w_r*recoverability with goal_match_floor=0.05 rumination guard (anchors with no payload OR goal_match below floor are invisible). Default pool: include_inactive=True, include_active=False. ValueError preconditions on use_anchor_sets=True AND use_sd039_anchor_payload=True. 6 sub-knobs surfaced through REEConfig.from_dims. V3-EXQ-496 5/5 PASS (UC1 module surface; UC2 master OFF no-op; UC3 6 admitted entries with max_priority 1.609, monotone-decreasing; UC4 Phase A goal-inactive all below floor / Phase B goal-active admitted with goal_match component dominant on top entry; UC5 components sum to priority within float epsilon).
  • MECH-293 waking ghost-goal probe search (hippocampal.awake_ghost_goal_probe_search) – read-side consumer of MECH-292. HippocampalModule.propose_trajectories() extended with a minority-budget ghost-seeded branch over the highest-priority bank entries’ anchor.z_world rather than the agent’s current z_world. Each ghost trajectory carries hypothesis_tag=True and metadata for downstream provenance routing; record_committed_trajectory strips the tag at commit boundary. mech293_ghost_fraction=0.2 default; mech293_replace_lowest_ranked=True preserves total candidate count. ARC-007 strict preserved – goal-match enters via MECH-292’s external ranking, not a hippocampal value head. ValueError precondition on use_mech292_ghost_bank=True. V3-EXQ-497 5/5 PASS (UC1 module surface; UC2 master OFF n_ghost=0; UC3 ghost branch fires n_ghost_admitted=4 max_priority=1.61 mean_goal_match_at_seed=0.998 reason=’ok’; UC4 hypothesis_tag preserved on every ghost + 28 value-flat candidates default-clean; UC5 budget-arithmetic clamp/cap/min-floor across 3 arms).
• MECH-163 V3 full-completion-gate substrate prerequisites cleared. All three substrate landings (SD-039 population layer, MECH-292, MECH-293) cleared 2026-04-27. V3-EXQ-495 (HABIT / PLANNED / ABLATED × A_DETOUR / B_NOVEL_CONTEXT × 7 seeds; THE V3-full-completion-gate metric: PLANNED-HABIT benefit-post-block gap >= 0.30 in detour, >= 4/7 seeds) is queued and is now the headline run; ~25h on Mac / ~40h on ree-cloud-1; machine_affinity=any.
• /diagnose-errors fixes 2026-04-27: (1) V3-EXQ-484 / 485 / 493 run_id naming-bug fix: source scripts emitted run_id ending in raw timestamp instead of _v3 suffix; sync_v3_results.py skipped them. Patched all three sources to construct run_id as f"{experiment_type}_{ts}_v3"; existing flat JSONs renamed + run_id field corrected; sync_v3 now picks them up cleanly; the 3 PASSes are now indexed pending awaiting next governance walk. (2) V3-EXQ-490 c1 gate-firing precondition root cause: VsRolloutGate’s hold trigger (V_s < 0.4) is unreachable under typical Phase 1 V_s dynamics because the identity-prediction proxy stays near 0.9-1.0 under aligned latents. Per user decision (smoke threshold-override path), queued V3-EXQ-490b (vs_gate_e1/e2_threshold=0.85, snapshot_refresh=0.95); claim_ids=[‘Q-040’] only (MECH-269b dropped because at smoke thresholds the gate fires regardless of stream staleness; skill rule 3 “err toward fewer tags”); supersedes V3-EXQ-490a. Q-040b (behavioural sufficiency) remains gated on Phase 2 forward-predictor V_s OR a substrate change wiring staleness_accumulator into VsRolloutGate.gate().
• Lit-pulls landed 2026-04-27: sequential sweep of 6 outstanding task_inbox lit-pulls (16 new entries across 6 claim directories) plus cowork-orchestrated 5-parallel lit-pull wave still in flight at snapshot time (EVB-0122 MECH-281, EVB-0123 Q-040, EVB-0124 SD-037, EVB-0125 MECH-057, EVB-0126 MECH-263). SD-033a A2/A3 lit-pull resolution brought MECH-262 lit_conf to 0.884; SD-033a lit_conf 0.87.
• Contracts suite: 183/183 contracts + 7/7 preflight PASS with all flags OFF after the 2026-04-27 wave (was 164/164 + 7/7 on 2026-04-26 before the wave). +12 MECH-293 contracts + remaining new SD-039 population / MECH-292 / MECH-293 contracts. Bit-identical-when-OFF guarantee preserved across the entire wave.
• Experiment count: 561 runner-side completions per runner_status.json 2026-04-27T08:04Z read (109 PASS / 242 FAIL / 66 ERROR / 144 UNKNOWN). +10 over the 2026-04-26 morning snapshot covering V3-EXQ-494 / 496 / 497 substrate-readiness PASSes plus V3-EXQ-484 / 485 / 493 PASS recovery after the run_id naming-bug fix.
• Pending review: 6 items per pending_review.md regenerated 2026-04-27T14:47:47Z (3 PASS + 3 runner-only ERROR/UNKNOWN/smoke for V3-EXQ-484/485/493). The 2026-04-27T14:11 governance cycle walked 9 indexed pending + 4 runner-only and applied: SD-039 / MECH-292 / MECH-293 substrate-readiness PASS clusters preserved as hold_pending_v3_substrate pending behavioural validation; V3-EXQ-433d SD-029 / MECH-256 reclassified non_contributory; V3-EXQ-418e SD-016 keeps does_not_support on path-1 div_weight=0.5; V3-EXQ-490 MECH-269b/Q-040 (×2 runs) reclassified non_contributory. Q-040 narrowed: split into Q-040a (precondition) and Q-040b (behavioural sufficiency). Substrate queue: SD-039 status flipped to implemented; MECH-292 + MECH-293 added as implemented; MECH-269b added as implemented_but_failing_validation; SD-016 received V3-EXQ-418e failure_record. Index rebuilt to 898 runs / 474 types. Next governance cycle gates on V3-EXQ-490b + V3-EXQ-495 outcomes.
• Queue (experiment_queue.json 2026-04-28): 2 items pending, both unclaimed.
  • V3-EXQ-495 pending – MECH-163 V3 full-completion gate; THE discriminative test for the VTA / hippocampally-planned arm of MECH-163 dual goal-directed systems. 3 conditions (HABIT value-flat proposals; PLANNED ghost-seeded proposals via MECH-293; ABLATED no goal anywhere) × 2 paradigms (A_DETOUR mid-episode blockage on the cached short corridor; B_NOVEL_CONTEXT cross-episode env swap to seed=137) × 7 seeds. P0 100ep encoder warmup + P1 100ep full pipeline + P2 50ep evaluation. Acceptance C1 (PLANNED ghost branch fires) + **C2 PLANNED-HABIT benefit-post-block gap >= 0.30 in detour,

    = 4/7 seeds (THE V3-full-completion criterion)** + C3 (HABIT >= ABLATED in standard episodes) + C4 (PLANNED.prox_r2 >= 0.7) + C5 (PLANNED.harm within 10% of HABIT). KL_PLANNED_HABIT first-step- action-distribution divergence recorded as diagnostic. Fishtank-viz per-step recording on FISHTANK_RECORD_SEED=42. machine_affinity=any; estimated_minutes=1500 (~25h Mac, ~40h ree-cloud-1).

  • V3-EXQ-490b pending – MECH-269b VsRolloutGate substrate- readiness probe; Q-040a precondition; supersedes V3-EXQ-490a. Smoke-only threshold override (vs_gate_e1/e2_threshold=0.85, snapshot_refresh=0.95) so the gate fires under typical Phase 1 V_s dynamics. PASS confirms substrate wiring (Q-040a precondition); Q-040b (behavioural sufficiency) stays gated on Phase 2 forward- predictor V_s or a substrate change. claim_ids=[‘Q-040’] only. machine_affinity=any; estimated_minutes=320.
• Current first-paper-gate bottleneck: V3-EXQ-495 is the headline V3-full-completion-gate run – all three substrate prerequisites cleared 2026-04-27, leaving only the runtime-budget decision. C2 PLANNED-HABIT benefit-post-block gap is THE gate metric. V3-EXQ-490b is the smaller upstream factorial for Q-040a (MECH-269b substrate- wiring precondition only). The EXQ-483 wired-but-inert pattern remains an open thread for the SD-037 / MECH-269b / MECH-295 cluster: V3-EXQ-484 / 485 / 493 all cleared as substrate-readiness PASSes on 2026-04-27 (post run_id naming-bug fix), validating SD-033a / SD-033b / MECH-295 substrate landings; behavioural recovery of approach_commit awaits the combined-cluster successor EXQ. Open promotion blockers documented in claims.yaml: MECH-294 within-cycle-vs-cross-cycle binding (Kay 2020 challenge); MECH-295 strong-vs-weak liking-bridge necessity (weak reading committed provisionally). SD-035 BLA hippocampal consumer wiring for retrieval_bias / remap_signal still deferred until V3-EXQ-474 behavioural signature confirmed.
• Substrate queue completeness pass 2026-04-27T18:20Z (post- reconcile): final state 52 queue items = 32 implemented + 3 implemented_but_failing_validation + 17 genuinely-pending/blocked, after a back-fill pass added 13 entries that had landed in substrate but were absent from evidence/planning/substrate_queue.json (SD-034, SD-035, MECH-266, MECH-272, MECH-273, MECH-275, MECH-279, MECH-284, MECH-285, MECH-287, MECH-288, MECH-290, MECH-295). All entries carry full schema (title, design_doc, depends_on_unresolved, unblocks_claims, implementation_hint, priority, implementation_status, implemented_utc, implemented_session, validation_experiment, metric_trajectory).

---

Status Snapshot (2026-04-27 — nightly docs sync, post-2026-04-26 substrate wave)

• SDs / MECHs moved to Implemented since the 2026-04-26 morning snapshot:
  • SD-039 substrate (hippocampal.anchor_goal_snapshot_payload) – dual-trace AnchorGoalPayload dataclass (z_goal_snapshot + wanting_strength + arousal_tag
    • last_vs + staleness_at_write + payload_written_step) + refresh-on-invalidate semantic preserving payload across mark_inactive + Anchor.goal_match cosine helper + AnchorSet.query_by_goal_match active+inactive dual-trace getter for the MECH-292 consumer. Module-level write-site population (from GoalState / VALENCE_WANTING / amygdala arousal tags) + V3-EXQ-494 falsifiable validation deferred to a follow-on session. 10/10 contract tests S1-S10 PASS.
  • SD-033b (pfc.ofc_analog) – OFC-analog as MECH-261 second consumer. Gate-modulated EMA state_code [1, state_dim] with eff_eta = update_eta * write_gate(“sd_033b”); zeroed-last-Linear bias head -> initial bias exactly zero; per-mode gate weights external_task=1.0 / internal_planning=0.5 / internal_replay=0.05 / offline_consolidation=0.3. Behavioural MECH-263 signatures (devaluation, same-sensory / different-task-role discrimination) deferred to env-extension EXQs. V3-EXQ-485 5-sub-test landing diagnostic smoke PASS.
  • MECH-269b (cortical_world_model.regional_verisimilitude_rollout_gating) – read-side consumer of MECH-269 Phase 1 per_stream_vs at the E1 _e1_tick site (before total_state cat / e1(…) call / extract_cue_context) and the per-tick E2_harm_a forward call site. Held substitution swaps current latent for snapshot when V_s[s] < per-side threshold (default 0.4 on both sides; refresh threshold 0.5; 0.4-0.5 dead-band Schmitt-trigger hysteresis). Precondition use_per_stream_vs=True (raises ValueError otherwise). Q-040 factorial validation queued as V3-EXQ-490.
  • MECH-295 weak-reading bridge (regulators.mech295_liking_bridge) – drive -> liking-stream -> approach_cue substrate. Two integration sites: (a) update_z_goal() writes anticipatory liking pulse to VALENCE_LIKING at the GOAL location (NOT current z_world), distinct from update_liking() consummatory; (b) select_action() reads per-candidate goal_proximity, computes drive*proximity, negates (E3 lower-is-better), composes additively with dacc_score_bias before e3.select(). Severed-bridge falsification arm at cue gain=0; weak-necessity reading committed provisionally. V3-EXQ-493 6-sub-test diagnostic (incl. UC5 SEVERED-BRIDGE COLLAPSE) smoke PASS.
• Architectural commitments registered 2026-04-26:
  • ARC-054 v4 -> v3 promotion – D_V trajectory selection promoted in rollout-horizon synaptic-EMA form (no TCL substrate dependency at V3); V4 form (phase-coherent V(t) integration via ARC-053 + MECH-225/226/228) remains v4-by-design. Design doc: docs/architecture/dv_temporal_depth_v3_form.md. V3-EXQ-491 validation queued.
  • MECH-271 V3 substrate plan – hypothesis tag as downstream routing committed for V3 in synaptic form (discrete routing table + audit hook for confabulation-vs-psychosis dissociation); V4 ephaptic-field-strength routing remains v4-by-design. Plan doc: docs/architecture/mech_271_routing_v3_substrate_plan.md. V3-EXQ-492 routing 4-arm queued behind the MECH-269b lock release.
  • V3/V4 phase substrate boundary memo added directly above this snapshot enumerating the architectural deferral table – ARC-053 / MECH-225 / MECH-226 / MECH-227 / MECH-228 / MECH-270 stay v4-by-design, the imaginary- plane / phase-channel deferral question is open, and the conditions under which a V4 promotion revisit triggers are documented.
• New claims registered 2026-04-26 (cingulate cluster lit-pull pass): MECH-294 (theta-burst multi-content packet; Kay 2020 cross-cycle alternation flagged as direct architectural challenge – explicit promotion blocker), MECH-269b (symmetric V_s gating; Q-040 factorial parent), MECH-295 (drive-amplified liking-stream as approach-cue bridge; weak provisional reading), Q-040 (V_s-generalisation-necessary-for-dACC question).
• Lit-pulls landed 2026-04-26: targeted_review_mech294theta_burst_packet (7 entries, sparse-but-not-falsifying), targeted_review_mech269b_vs_rollout gating (7 entries, mean conf 0.69; symmetric-application novelty flagged), targeted_review_mech295_liking_approach_bridge (6 entries, mean conf 0.77; strong-vs-weak necessity flagged for user resolution). Plus 5 MECH-280 literature_evidence/v1 entries written into the existing targeted_review_sd_037_orexin_kinetics/ folder (de Araujo Salgado 2023 Neuron, Marino 2020 PNAS, Johnson 2012 Prog Brain Res, Sakurai 2014 NRN, Mileykovskiy 2005 Neuron); MECH-280 lit_conf 0 -> 0.878. Plus targeted_review_ghost_goal_search (7 entries) – the seed for SD-039 / MECH-292 / MECH-293 / ARC-060 registration earlier in the day.
• Contracts suite: 164/164 contracts + 7/7 preflight PASS with all flags OFF after the 2026-04-26 wave (was 150/150 + 7/7 on 2026-04-25 before the wave). Bit-identical-when-OFF guarantee preserved across SD-039 substrate
  • SD-033b + MECH-269b + MECH-295.
• Experiment count: 551 runner-side completions per runner_status.json 2026-04-26 read (109 PASS / 241 FAIL / 66 ERROR / 135 UNKNOWN; v3 subset 93 PASS / 228 FAIL / 66 ERROR / 135 UNKNOWN). +1 over the 2026-04-26 morning snapshot.
• Pending review: 0 items as of pending_review.md regenerated 2026-04-26T16:19:34Z (down from 1 on the 2026-04-26 morning snapshot). The 2026-04-26T15:39 governance cycle reclassified V3-EXQ-483a manifest supports -> non_contributory per-claim (SD-037 / MECH-280 / MECH-281 remain hold_pending_v3_substrate); same wired-but-inert pattern as the V3-EXQ-471 / 478 / 480 cluster. 9 hold_pending_v3_substrate decisions applied via apply_decision_batch.py: ARC-051, ARC-060, MECH-269b, MECH-291, MECH-292, MECH-293, MECH-294, MECH-295, SD-039.
• Queue (experiment_queue.json 2026-04-27): 6 items.
  • V3-EXQ-433d in flight – SD-029 / MECH-256 event-conditioned comparator with the EXQ-479 calibrated curriculum (interval=10, num_hazards=2, hazard_harm=0.02, adjacent_only=True); STEPS_PER_EP bumped 120 -> 200; supersedes V3-EXQ-433c. Auto-claimed by DLAPTOP-4.local 2026-04-26T15:01Z.
  • V3-EXQ-418e pending – SD-016 Path 1 4-arm A0_off / A1_writes_only / A2_div_only / A3_writes_plus_div ablation; supersedes V3-EXQ-418d.
  • V3-EXQ-484 pending – SD-033a distractor-resistance under MECH-261 internal_replay gate; 3-arm deterministic at the SalienceCoordinator + LateralPFCAnalog interface (no agent loop). Smoke PASS 2026-04-26.
  • V3-EXQ-485 pending – SD-033b OFC-analog landing diagnostic; 5 sub-tests paralleling EXQ-456. Smoke PASS 2026-04-26.
  • V3-EXQ-490 pending – MECH-269b symmetric V_s gating substrate-readiness diagnostic; Q-040 factorial ON_OFF vs ON_ON arms with use_broadcast_override + use_dacc + drive_weight=2.0 + full V_s invalidation circuit + use_vs_commit_release ON; only manipulated variable is use_vs_rollout_gating. ~50-55 min/arm.
  • V3-EXQ-493 pending – MECH-295 weak-reading bridge validation; 6 sub-tests including UC5 SEVERED-BRIDGE COLLAPSE falsifiable signature. Smoke PASS 2026-04-26.
• Current first-paper-gate bottleneck: V3-EXQ-490 + V3-EXQ-493 jointly dissect the EXQ-483 wired-but-inert pattern. Q-040 factorial dispatches: V3-EXQ-490 PASS (gate fires AND approach_commit recovery AND non-zero dacc_score_bias) -> cortical-side V_s gating dominates and SD-037 reopens. V3-EXQ-490 FAIL on C2/C3 with C1 PASS -> evidence redirects at MECH-295 as the dominant blocker. V3-EXQ-493 separately validates the liking-bridge mechanism + the severed-bridge collapse falsification. Both PASS -> joint contribution and a combined-cluster behavioural EXQ follows. V3-EXQ-433d gates the SD-003 successor track; V3-EXQ-418e gates the SD-016 cue_action_proj forward-path re-validation. Open promotion blockers documented in claims.yaml: MECH-294 within-cycle-vs-cross-cycle binding (Kay 2020 challenge); MECH-295 strong-vs-weak liking-bridge necessity (weak reading committed provisionally). SD-035 BLA hippocampal consumer wiring for retrieval_bias / remap_signal still deferred until V3-EXQ-474 behavioural signature confirmed.
• Governance: Mid-day governance cycle 2026-04-26T15:39 walked 1 indexed pending (V3-EXQ-483a) and applied the 9 hold_pending_v3_substrate decisions listed above. SD-037 substrate_queue refreshed with implementation_status implemented_but_failing_validation and a cross-pointer to MECH-295 as the likely root cause; SD-039 added as a new ready substrate-queue entry (ghost-goal anchor payload). Predecessor V3-EXQ-483 manifest reclassified non_contributory -> superseded. Index rebuilt to 889 runs / 468 types. Next governance cycle gates on V3-EXQ-490 + V3-EXQ-493 outcomes.

---

V3/V4 Phase Substrate Boundary – Architectural Commitment (2026-04-26)

The V3 working-model uses a synaptic approximation of regional verisimilitude (V_s), its temporal-depth integration (D_V), its routing realisation (MECH-271), and its fast- broadcast invalidation (MECH-287). The full phase-channel substrate – ARC-053 Temporal Coherence Loop, MECH-225 oscillatory cross-frequency multiplexing, MECH-226 TCL biophysical substrate (inferior olive + cerebellum + thalamus + cortex), MECH-228 ephaptic field-level coherence support, MECH-227 anaesthesia-collapse model, MECH-270 ephaptic substrate of V_s – stays v4-deferred by architectural commitment, not by substrate prerequisite.

The architectural bet: the synaptic V3 form is sufficient for V3 working-model deliverables (closed-loop agent with V_s invalidation runtime, D_V-aware rollout selection, hypothesis- tag-as-routing). The V4 phase substrate refines, not replaces, the V3 form.

Two v4-held claims have V3 forms landing now:

• ARC-054 (D_V trajectory selection) – V4 form (phase-coherent V(t) integration) remains v4. V3 form (rollout-horizon synaptic EMA over V_s readout) promoted to v3 2026-04-26. Design doc: docs/architecture/dv_temporal_depth_v3_form.md.
• MECH-271 (hypothesis tag as downstream routing) – V4 form (phase-channel routing via ephaptic field strength) remains v4. V3 substrate landing plan committed 2026-04-26: discrete routing table + audit hook for confabulation-vs-psychosis dissociation. Plan doc: docs/architecture/mech_271_routing_v3_substrate_plan.md.

Full V4 deferral table, what the synaptic forms cover and don’t cover, the imaginary- plane (phase-channel) deferral question, and the conditions under which a V4 promotion revisit triggers: docs/architecture/v3_v4_phase_substrate_boundary.md.

Governance hook: the v4-held claims currently produce hold_pending_v3_substrate recommendations, which is incorrect labelling for these specifically – they are v4-by-design, not v4-by-prerequisite. A separate governance-tooling session should add a held_v4_by_architectural_commitment recommendation type that reads from the boundary doc.

---

Status Snapshot (2026-04-26 — nightly docs sync, post-2026-04-25 substrate wave)

• SDs moved to Implemented since the 2026-04-25 snapshot:
  • SD-037 (regulators.broadcast_override) – orexin/hypocretin-analog BroadcastOverrideRegulator. Scalar override_signal in [0,1] driven by SD-012 drive_level + sustained-threat rolling-window magnitude over z_harm, EMA-smoothed. Consumed at three sites: PAG freeze-gate exit_threshold scaled by (1 + alpha_override * override_signal); SalienceCoordinator update_signal("override_signal", ...) biases operating-mode aggregate toward external_task; GoalState seeding amplified effective_drive *= (1 + (override_goal_seeding_gain - 1) * override_signal). MECH-094 simulation_mode gate: tick(simulation_mode= True) returns cached signal unchanged. Failure-mode predictions: PWS-hyperphagia analog (saturated override -> >=2x approach-commit rate); narcolepsy/cataplexy analog (lost override -> <30% approach- commit); catatonic lock-in escape (raises PAG exit_threshold under sustained drive+harm).
  • Sleep Aggregation Cluster Phase A (SleepLoopManager scaffolding) – wraps the existing SD-017 surface (run_sleep_cycle / enter_sws_mode / run_sws_schema_pass / enter_rem_mode / run_rem_attribution_pass / exit_sleep_mode). New SleepPhase enum (6 phases: WAKING / SLEEP_ENTRY / SWS_ANALOG / PHASE_SWITCH / REM_ANALOG / WRITEBACK) + SleepCycleState dataclass. Master flag use_sleep_loop (default False) + sleep_loop_episodes_K (default 1) + sleep_loop_require_ passes (default True). notify_episode_end() hooked at start of REEAgent.reset() so sleep operates on the final waking state.
  • MECH-285 Sleep Phase B (sleep.replay_sampler) – SleepReplaySampler. At SLEEP_ENTRY freezes StalenessAccumulator.snapshot(), then draws N seeds from AnchorSet.all_with_dual_trace() (active + inactive, Bouton 2004 dual-trace preserved) with softmax(staleness/temperature) priority. Stateless within cycle; uniform-fallback when no accumulator. Phase B is no-op consumer – draws land in mech285_* metrics only.
  • MECH-272 Sleep Phase C (sleep.routing_gate) – RoutingGate. State- conditioned channel weights {anchor_channel, probe_channel} flipping across SWS_ANALOG / REM_ANALOG / WAKING rows per the design-doc table. Per-draw RoutedEvents surfaced as mech272_* cycle metrics.
  • MECH-275 Sleep Phase D (sleep.bayesian_aggregator) – BayesianAggregator. Per-domain per-region Gaussian posteriors over residuals; conjugate mean-and-variance update gated by RoutedEvent.probe_channel * probe_gain (probe<=0 skipped, counted as mech275_n_skipped_zero_ probe). Snapshot+decay contract: snapshot() deep-copies live posteriors at PHASE_SWITCH (frozen pre-REM); decay_factor multiplies live variance per cycle. Place-domain default with (scale, segment_ id) region key matching MECH-284.
  • MECH-273 Sleep Phase E (sleep.self_model_writeback) – SelfModel Aggregator subclass of MECH-275 specialised on SD-003 causal_sig posterior. offline_gradient_pass(e2_harm_s, replayed_regions, n_steps, domain='self', use_snapshot=True) reads posterior means from last_snapshot (SWS-only frozen copy at PHASE_SWITCH); constructs synthetic batch at E2_harm_s input dims; trains via Adam at waking_lr * offline_lr_scale for n_steps bounded MSE steps. MECH-094 exception scoped: optimiser constructed locally over e2_harm_s.parameters() only – no other module’s params touched. NEW API: StalenessAccumulator.partial_decay(replayed_regions, decay_factor=0.5) multiplicatively decays only supplied region keys (clamped, drops below drop_epsilon).
  • SD-016 Path 1 (e1.context_memory_diversification_loss) – auxiliary mean-squared-off-diagonal-cosine loss on normalised ContextMemory slot vectors. Provides gradient pressure for slot symmetry-breaking missing in EXQ-418d 4-arm writepath ablation (FAILed across all arms with attn_entropy_mean ~2.76 near uniform reference 2.7726 and bimodal seed pattern: seed 42 ~0.46 div, seeds 43/44 collapse <1e-4). Smoke verified slot_div climbs 0.2->0.5->1.0 across arms; new sd016_diversification_weight config wired through E1Config + REEConfig.from_dims (default 0.0 – backward compatible).
• Contracts suite: 150/150 PASS (143 contracts + 7 preflight) with all flags OFF after the 2026-04-25 wave (was 91/91 contracts + 7 preflight before the wave). Bit-identical-when-OFF guarantee preserved across SD-037 + 5 sleep phases + SD-016 Path 1.
• Experiment count: 550 runner-side completions per runner_status.json 2026-04-26 read (108 PASS / 241 FAIL / 66 ERROR / 135 UNKNOWN; v3 subset 92 PASS / 228 FAIL / 66 ERROR / 135 UNKNOWN). +5 completions over the 2026-04-25 snapshot.
• Pending review: 1 item as of pending_review.md regenerated 2026-04-26T00:44:10Z (down from 13 on the 2026-04-25 snapshot). The remaining item is V3-EXQ-483 FAIL whose successor V3-EXQ-483a is in flight.
• Queue (experiment_queue.json 2026-04-26): 2 items.
  • V3-EXQ-483a in flight – SD-037 broadcast-override 4-arm with WARMUP_EPISODES=200 + substrate-readiness fallback acceptance; supersedes V3-EXQ-483. Auto-claimed by DLAPTOP-4.local 2026-04-25T23:29Z. EXQ-483 confirmed substrate-readiness (override 0.0 -> 0.56 mean / 0.62 max in ON arms; PAG releases 5.3 -> 9.0-9.3) but behavioural metrics were uninterpretable because approach_commit=0.0 in ALL arms including the SD-036-only baseline. Two changes: (1) WARMUP_EPISODES 60 -> 200 to give baseline arm enough exposure to potentially produce non-zero approach behaviour; (2) acceptance logic adds substrate- readiness fallback path used only when baseline arm yields no behavioural signal (override_mean > 0.30 AND PAG release ratio ON_ON/ON_OFF > 1.30). Behavioural path remains preferred when baseline is non-zero.
  • V3-EXQ-418e pending – SD-016 Path 1 4-arm A0_off / A1_writes_only / A2_div_only / A3_writes_plus_div ablation; supersedes V3-EXQ-418d. Acceptance C1 attn_entropy<2.65 AND C2 div>0.10 ALL 3 SEEDS AND C3 behavioural delta >=0.20 AND C4 A1 replicates 418d FAIL. C2 raised from 2/3 to 3/3 because 418d showed bimodal seed pattern – substrate-level fix must escape collapse on every seed.
• Current first-paper-gate bottleneck: V3-EXQ-483a SD-037 validation (in flight) gates the orexin-analog failure-mode signature – PWS- hyperphagia (saturated override -> >=2x approach-commit), narcolepsy/ cataplexy (lost override -> <30% approach-commit), and the V3-EXQ-471 catatonic-lock-in escape signature. V3-EXQ-418e SD-016 Path 1 validation is the parallel SD-016 cue_action_proj design-rethink resolution path – if Path 1 lifts the bimodal seed pattern across all 3 seeds, the diversification-loss substrate is the answer; if it does not, deeper ContextMemory write-architecture redesign is required. SD-032 cluster behavioural follow-through remains the primary cingulate-track blocker (V3-EXQ-445a/b/c FAILed; V3-EXQ-325d FAILed with zero between-arm gradient on the SD-032c AIC-analog descending-modulation arm); the cluster awaits the next governance cycle pass.
• Governance: No governance cycle run in this nightly window. Next cycle should ingest V3-EXQ-483a + V3-EXQ-418e outcomes, the SD-035 / MECH-266 landings, and the 2026-04-24 + 2026-04-25 substrate waves.

---

Status Snapshot (2026-04-25 — nightly docs sync, post-Phase-3 wave)

• SDs moved to Implemented since the 2026-04-24 snapshot:
  • MECH-284 Phase 3 (hippocampal.staleness_accumulator) – region-indexed staleness with per-tick leak (default leak_factor=0.995), attribution_mode equal / stream_overlap, staleness_clip=1.0, lookup_by_anchor_key getter consumed by MECH-269 online hysteresis. Integration site: HippocampalModule.tick_anchor_set peek-not-drains the _broadcast_event_queue so MECH-287 broadcast events propagate transparently.
  • MECH-269 online hysteresis swap (AnchorSet.tick_hysteresis accepts optional staleness_lookup; orthogonal flag use_mech284_hysteresis, default OFF). With both flags ON, V_s_anchor = V_s(r) - staleness[r] drives anchor-reset; default OFF preserves the Phase 2 internal proxy so the substrate is non-invasive.
  • MECH-290 (hippocampal.backward_trajectory_credit_sweep) – Foster & Wilson 2006 reverse replay. record_committed_trajectory at BetaGate elevation; backward_credit_sweep at completion-signal release; per-step credit = outcome_quality * gamma^(T-1-t) -> ResidueField.update_valence(VALENCE_WANTING). Reset on episode boundary.
• Contracts suite: 91/91 PASS with all flags OFF (was 85/85 before the Phase 3 wave); preflight 7/7 PASS; bit-identical to the pre-Phase-3 HEAD with master switches off. Activation smokes – ARM0/1/2 for MECH-284 and end-to-end direct-wiring tests for MECH-290 – all PASS 2026-04-24.
• Experiment count: 545 runner-side completions per runner_status.json 2026-04-25 read (108 PASS / 239 FAIL / 66 ERROR / 132 UNKNOWN; v3 subset 88 PASS / 226 FAIL / 64 ERROR / 132 UNKNOWN). Indexer carries 881 indexed runs as of the 2026-04-24 cowork rebuild; indexer-vs-runner gap is the historical pre-runner_status archive plus per-seed manifests collapsed to single queue entries. Next indexer rebuild is gated on V3-EXQ-478 returning.
• Pending review: 13 items as of pending_review.md regenerated 2026-04-24T11:54:11Z (down from 25 on the 2026-04-23 snapshot; cowork-a wave reviewed V3-EXQ-433c / 449b / 447a and resolved several unknowns).
• Queue: EMPTY (experiment_queue.json items: [] as of 2026-04-25T01:14Z). Active waterfall:
  • V3-EXQ-478 in flight – MECH-284 Phase 3 validation diagnostic, OFF vs ON x 2 seeds; metrics freeze_recommit_count, anchor_reset_count, mean_staleness_peak, action_class_entropy. Auto-claimed by DLAPTOP-4.local 2026-04-24T13:22Z; runner_status carries it as UNKNOWN pending completion. PASS unlocks the previously gated V3-EXQ-445d / 449c / 455a / 476 cascade. FAIL forces a Phase 3 redesign before the cascade proceeds.
  • V3-EXQ-479 next-up – SD-029 fix2 superseding 470a; queued by cowork-2026-04-24-a with the curriculum / agent-caused-elicitation correction.
  • V3-EXQ-476 / 476a / 476b returned ERROR/UNKNOWN before MECH-284 Phase 3 was available; 476a and 476b queued in discussed_experiment_dirs for review.
  • V3-EXQ-449c / 445d / 455a all errored under the pre-MECH-284 substrate and await V3-EXQ-478 PASS before re-queueing.
  • V3-EXQ-418c remains needs_user_review (SD-016 cue_action_proj design rethink open, anchored on the 2026-04-24 V3-EXQ-477 key_proj.bias dominance diagnosis).
  • V3-EXQ-137 FAILed 2026-04-24T02:21Z (MECH-097 PPS commit locus, instrumentation-fix); V3-EXQ-477 FAILed 2026-04-24T08:06Z (SD-016 ContextMemory slot-store / attention-uniformity diagnostic).
• Current first-paper-gate bottleneck: V3-EXQ-478 (MECH-284 Phase 3 validation, in flight) gates the V_s-gated cascade. SD-032 cluster behavioural follow-through remains the primary cingulate-track blocker (V3-EXQ-445a / 445b / 445c all FAILed; V3-EXQ-325d FAILed with zero between-arm gradient on the SD-032c AIC-analog descending-modulation falsification signature; V3-EXQ-454 FAILed on ARC-016 adaptive commitment_threshold). SD-016 cue_action_proj forward-path is now in the design-rethink anchor (V3-EXQ-477 FAIL diagnosed key_proj.bias dominance as the substrate-level problem), gating V3-EXQ-418c. Governance-cycle pass remains pending for the SD-032 behavioural FAILs, the SD-035 / MECH-266 landings, the V_s invalidation runtime substrate landings, and the 2026-04-24 Phase 3 wave.

---

Status Snapshot (2026-04-24 — nightly docs sync, queue refresh)

• No new SDs or governance decisions this session (nightly docs-only sync following experiment queue refresh by PM session 2026-04-23).
• Queue refresh: queue grew from the 2026-04-23 PM snapshot’s “3 claimed items” (V3-EXQ-447 / 451 / 445a – all since cleared) to 6 items (1 claimed, 5 pending):
  • V3-EXQ-476 (pending, priority 70, diagnostic) – MECH-269 V_s validation entropy probe, cascade gate for the V_s-gated cascade track (EXQ-445d / EXQ-449c / EXQ-455a). Baseline agent + V_s flags ON vs OFF; measure action_class_entropy. PASS = ON entropy > OFF entropy by >=0.1 in >=2/2 seeds. Queued 2026-04-24 – this is the end-to-end validation item that the 2026-04-23 snapshot flagged as “planned but not yet queued” for the 2026-04-22 V_s invalidation runtime substrate wave. FAIL/INCONCLUSIVE means MECH-284 Phase 3 consumer must land before downstream cascade can run.
  • V3-EXQ-449c (pending, priority 50, evidence, 150 min) – MECH-074b BLA retrieval bias V_s-gated ablation; depends_on: V3-EXQ-445d. PASS = action_class_entropy ON - OFF >= 0.1 AND harm_rate reduced in >=2/3 seeds.
  • V3-EXQ-433c (claimed DLAPTOP-4.local 2026-04-23T23:23:48Z, priority 55, evidence, 90 min) – SD-029 event-conditioned MECH-256 comparator with curriculum ON + scripted agent-caused elicitation; supersedes V3-EXQ-433b. Fix: SD-029 curriculum enabled in P0 / P1 / eval (scheduled_external_hazard_enabled=True, interval=25, prob=1.0, adjacent_only=False); deterministic move onto an adjacent hazard when trials_collected[‘agent_caused_hazard’] is short; C0 sufficiency gate (n_agent / n_env >= 20 in >=3/4 seeds). If C0 fails, outcome=FAIL but per-claim evidence_direction=’inconclusive_insufficient_events’ (not ‘weakens’) so governance scores are not corrupted by a trials-shortage run. Re-opens the SD-003 successor track.
  • V3-EXQ-449b (pending, priority 52, diagnostic, 30 min) – SD-016 cue_action_proj consumer fix verification (z_world residual concat); supersedes V3-EXQ-449a. EXQ-449a localised the collapse to a uniform-attention bottleneck inside extract_cue_context (ContextMemory slots init at randn*0.01 so key_proj’s bias dominates; all keys look identical; softmax = uniform entropy 2.7726; bmm(uniform, v) constant across batch -> cue_context constant -> cue_action_proj output had per-channel std ~2.7e-8 in g2). Fix (predictors/e1_deep.py): cue_action_proj input changed from cue_context alone (latent_dim=64) to [cue_context, z_world] (concat, latent_dim+world_dim=96). cue_terrain_proj left unchanged. Same three-regime protocol as 449a (g1 supervised-active, g2 frozen, g3 detach-bypassed); acceptance pivots from “find the offender” to “verify the offender is gone”. Smoke-test 2026-04-23 dry-run (P0=2 P1=3 eval=4 ep): g2 per-channel std = 2.957e-3, primary_pass=True. Unblocks V3-EXQ-418c.
  • V3-EXQ-418c (pending, priority 50, evidence, 60 min) – SD-016+SD-017 context-conditioned action with cue_action_proj consumer fix active; supersedes V3-EXQ-418b. Re-run of EXQ-418a using the SAME script (fix is upstream in e1_deep.py and activates automatically when sd016_enabled=True, which 418a already sets). EXQ-418/418a/418b all FAILed three times with action_bias_divergence=0.0 under the broken substrate.
  • V3-EXQ-137 (pending, priority 40, evidence, 180 min) – MECH-097 PPS commit locus: PPS_LOCUS_ON vs ABLATED, backlog EVB-0137. Instrumentation fixed 2026-04-24 (verdict print, outcome field, timestamp_utc, EXPERIMENT_PURPOSE); smoke-test PASS.
• Experiment count: 844 runs (indexer rebuilt 2026-04-23 PM; runner_status.json last_updated 2026-04-23T20:23:18Z, 527 runner-side completions: 235 FAIL / 108 PASS / 62 ERROR / 122 UNKNOWN; no new completions in runner_status.json since the 2026-04-23 snapshot). Next indexer rebuild will refresh once V3-EXQ-476 and the other newly-queued runs return.
• Pending review: 25 items as of pending_review.md regenerated 2026-04-23T17:49:07Z (24 PASS, 0 FAIL, 1 UNKNOWN for V3-EXQ-471). PASS queue is dominated by the SD-033 cluster landings (EXQ-456, 460, 462-468) across multiple timestamps; the UNKNOWN clears on the next indexer rebuild.
• Current first-paper-gate bottleneck: V_s invalidation runtime end-to-end validation is now the next gate. V3-EXQ-476 cascade-gate entropy probe is queued; PASS unlocks V3-EXQ-449c and the downstream V_s-gated cascade (EXQ-445d / 455a). SD-032 cluster behavioural follow-through remains the primary cingulate-track blocker (V3-EXQ-445a / 445b / 445c all FAILed; V3-EXQ-325d FAILed with zero between-arm gradient on the SD-032c AIC-analog descending-modulation falsification signature; V3-EXQ-454 FAILed on ARC-016 adaptive commitment_threshold). The SD-003 successor track is re-opened by V3-EXQ-433c now that the agent_caused_hazard r2=0.0 from V3-EXQ-433b has been diagnosed as a curriculum-sufficiency issue (0 agent-caused trials collected in every seed because the env relied on organic elicitation) rather than a MECH-256 architectural failure. SD-016 cue_action_proj forward-path is now unblocked by the V3-EXQ-449b verification (2026-04-23 dry-run g2 per-channel std = 2.957e-3, primary_pass=True), which in turn re-enables V3-EXQ-418c.

---

Status Snapshot (2026-04-23 PM — lit-pull + docs sync)

• Literature pull completed: MECH-074a/c/d (PM session 2026-04-23T11:24Z–11:34Z). Six new entries added to evidence/literature/targeted_review_connectome_mech_074/:
  • MECH-074a (3 entries): Paré 2003 (connectome_mechanistic_review, conf=0.80); Roozendaal et al. 1999 PNAS (behavioral_animal, conf=0.82 — direct beta-adrenergic gate evidence); Bass & Manns 2015 (electrophysiology_lfp, conf=0.72 — BLA stimulation → CA3-CA1 gamma synchrony → STDP circuit mechanism).
  • MECH-074c (2 entries): Ciocchi et al. 2010 Nature (electrophysiology_single_unit, conf=0.78 — CeL required for fear acquisition, CeM drives output, GABAergic disinhibition); Walker & Davis 2008 (connectome_mechanistic_review, conf=0.75 — CeA(M) rapid phasic output vs BNST sustained fear, directly grounds the fast_prime vs MECH-046 distinction).
  • MECH-074d (1 entry): Redondo et al. 2014 Nature (behavioral_animal, conf=0.62, mixed — DG engram can switch valence, BLA engram cannot; supports BLA attribution stability but does not directly test PE-triggered partial remap amplitude). claim_evidence.v1.json updated: MECH-074a lit_conf=0.840, MECH-074c lit_conf=0.782, MECH-074d lit_conf=0.560. Index rebuilt: 923 literature entries across 443 experiment types.
• Experiment count: 844 runs (indexer rebuilt 2026-04-23 PM; prior session spec showed stale count of 525 — now corrected in spec and README).
• Pending review: 0 indexed pending; 10 runner-only UNKNOWN (V3-EXQ-456, 460, 462–468, 471, 447 — all from V_s invalidation runtime / SD-032 cluster landings; will clear after next indexer rebuild with those results indexed).
• No new SDs or governance decisions this session (docs-only sync pass following the PM lit-pull).

---

Status Snapshot (2026-04-23)

• V_s invalidation runtime substrate wave LANDED 2026-04-22. Six substrates landed in a coordinated 2026-04-22 session implementing the architecture doc REE_assembly/docs/architecture/v_s_invalidation_runtime.md:
  • SD-036 GABAergic cross-stream decay regulator (ree_core/regulators/gabaergic_decay.py) – broadly-projecting tonic decay applied out-of-place across registered latent streams (z_harm tau=0.05, z_harm_a tau=0.02, z_beta tau=0.03 by default; drive accumulator intentionally excluded). Global gaba_tone multiplier in [0, 2] models benzo-analog (>1) and withdrawal / chronic-stress analog (<1). Wired in agent.sense() right after LatentStack.encode() so all downstream consumers see the decayed latent on the same tick.
  • MECH-279 PAG freeze-gate (ree_core/pag/freeze_gate.py) – committed- freeze substrate keyed on duration_above_threshold * z_harm_a > theta_freeze (default 2.0). Exit threshold scales with SD-036 gaba_tone, so the same GABAergic system gates BOTH freeze entry AND freeze exit (architectural prediction: GABA agonists treat freeze catatonia). Action-class no-op injection during freeze; simulation_mode gated.
  • MECH-269 base / Phase 1 per-stream V_s (ree_core/hippocampal/module.py::update_per_stream_vs) – foundation observable: identity-prediction proxy EMA over registered streams (z_world / z_self / z_harm_s / z_harm_a / z_goal / z_beta); seeds at 1.0 on first tick, drops on latent change. Forward-predictor routing (SD-007 reafference for z_world, SD-011 harm forward for z_harm_s) reserved for Phase 2 consumer wiring.
  • MECH-288 event segmenter Phase 2 (ree_core/hippocampal/event_segmenter.py) – two-scale boundary detector: fast pe_threshold on (z_world, z_self) (window=200, threshold=0.65) + slow BOCPD-Gaussian on (z_goal,) (hazard=1/40, posterior_threshold=0.5, top-k=20). Emits BoundaryEvents with nested outer.inner segment IDs; slow fire forces outer+=1, inner=0 and suppresses same-tick fast; force_boundary API for scripted injection. BOCPD uses underflow-robust Adams & MacKay 2007 recursion with Welford online variance.
  • MECH-287 invalidation trigger Phase 2 iv (ree_core/regulators/invalidation_trigger.py) – BoundaryEvent subscriber re-emitting graded BroadcastEvents (strength = posterior * gain; NO binary thresholding of strength). Phasic/tonic guardrail (Aston-Jones & Cohen 2005; Clewett 2025 failure signature 2) via rolling mean over tonic_window=50 past-tick posteriors; suppresses phasic broadcast when tonic estimate exceeds tonic_threshold=0.5. Verdict-3 architectural commitment (V_s foundation lit-pull synthesis): the trigger IS the subscriber, not an independent CA1/CA3 mismatch comparator stage – the biological two-stage loop is collapsed to a subscription on the MECH-288 boundary queue. Dissociation contract C5 verifies: lesioning the segmenter silences the trigger regardless of its internal tonic state.
  • MECH-269 Phase 2 ii AnchorSet (ree_core/hippocampal/anchor_set.py) – scale-tagged anchor store keyed on (scale, segment_id, stream_mixture). Dual-trace preservation per Bouton 2004: remap on the same (scale, stream_mixture) marks the outgoing anchor INACTIVE and retains it in all_anchors(); never erased. k=5 consecutive-below-threshold hysteresis on V_s_anchor = avg(V_s over mixture) - staleness (staleness monotonic in tick - last_accessed). FIFO soft-cap at 128 active anchors per scale. BoundaryEvent consumer via tick_anchor_set(latent_state, events); Phase 2 stream_mixture stand-in is tuple(sorted(per_stream_vs.keys())) at anchor-creation tick (learned attribution-head version reserved for Phase 3 MECH-284).
  • MECH-269 Phase 2 iii T4 per-region V_s (extended module.py) – promotes flat per_stream_vs[stream] -> float to per_region_vs[(scale, segment_id)][stream] -> float keyed on active AnchorSet regions. Two reset paths: (1) passive hysteresis via tick_anchor_set marking anchors inactive; (2) explicit via apply_invalidation_broadcasts_to_regions(broadcasts) dropping region V_s and mark_inactive’ing the matching anchor on MECH-287 BroadcastEvents (keyed on source_scale, source_segment_id_old). Peek-not-drain on the broadcast queue preserves events for Phase 3 MECH-284 staleness accumulator consumer. Idempotent on repeated broadcasts. All six landed via 85/85 contracts + 7/7 preflight PASS with flags OFF (bit-identical to legacy), plus dedicated contract tests for each phase: MECH-269 Phase 1 (5 tests), MECH-288 (7 tests), MECH-287 (5 tests incl. verdict-3 dissociation C5), MECH-269 Phase 2 ii (9 tests incl. 2 integration smokes for agent-level flag behaviour), MECH-269 Phase 2 iii T4 (6 tests incl. 1 integration smoke). Activation smokes confirmed expected signatures: CeA synthetic threat -> graded mode_prior/fast_prime; BLA synthetic arousal -> inverted-U cap; BLA synthetic PE-spike -> Moita 2004 remap; default agent
  • Phase 1 flag ON seeds per_stream_vs at 1.0 and drops on perturbation; forced fast boundary installs anchor under segment_id “0.1” and populates per-region V_s under the new region key. Design doc: docs/architecture/v_s_invalidation_runtime.md; anchor-selection doc: docs/architecture/hippocampal_anchor_selection.md. End-to-end combined- cluster validation (V3-EXQ-476: matched re-run of EXQ-475 with the full circuit on) is planned; not yet queued as of 2026-04-23.
• Experiments: 525 total completions unchanged since 2026-04-22 snapshot (runner_status.json last_updated 2026-04-22T01:11Z) – the V_s invalidation runtime wave landed via contract tests + activation smokes only. PASS/FAIL breakdown: 108 PASS / 234 FAIL / 62 ERROR / 121 UNKNOWN. Next indexer rebuild will refresh after V3-EXQ-476 and the pending SD-032 / SD-003 successor results land.
• Pending review: 10 items as of 2026-04-22T23:12:38Z (down from 46 at the 2026-04-21T19:54Z snapshot). All 10 are runner-UNKNOWN because the index is stale ahead of the next indexer rebuild (python evidence/experiments/ scripts/build_experiment_indexes.py from REE_assembly/ root clears them). Queue IDs: V3-EXQ-456 / 460 / 462 / 463 / 464 / 465 / 466 / 467 / 468 / 471.
• Current first-paper-gate bottleneck: SD-032 cluster behavioural follow- through remains the primary gate. V3-EXQ-445a (SD-032b dACC full-pipeline fix for the EXQ-445 monostrategy collapse + terrain-prior inversion, claimed EWIN-PC) is the decisive test still in flight; V3-EXQ-445b / 445c have both FAILed and V3-EXQ-325d FAILed on the SD-032c descending- modulation falsification signature. V3-EXQ-447 (SD-032d deterministic validation, claimed ree-cloud-2) and V3-EXQ-451 (Q-034 hazard/resource threshold retest, claimed EWIN-PC) are the remaining two claimed experiments. SD-003 successor track (V3-EXQ-433a MECH-256/SD-029 FAIL, V3-EXQ-452 MECH-257 dual-function E2 diagnostic FAIL) and ARC-007 path- memory track (V3-EXQ-397c) remain alive. Secondary bottleneck: V_s invalidation runtime end-to-end validation deferred – V3-EXQ-476 (combined-cluster re-run of EXQ-475 with the full circuit on) is planned but not yet queued.

---

Status Snapshot (2026-04-22)

• SD-035 amygdala analogue LANDED 2026-04-21. BLA + CeA peer modules (ree_core/amygdala/bla.py, ree_core/amygdala/cea.py) non-trainable arithmetic. BLA instantiates MECH-074a inverted-U encoding_gain (Roozendaal 2011), MECH-074b content-selective retrieval_bias (LaBar & Cabeza 2006 per-trace weight vector, not scalar), and MECH-074d attribution-gated PE-spike remap (Moita 2004). CeA instantiates MECH-046 pre-softmax mode-prior (LeDoux 1996 “low road” / Pessoa & Adolphs 2010, distinct from SD-032c AIC which biases mode-SWITCH threshold rather than mode SELECTION) and MECH-074c fast_prime (Mendez-Bertolo 2016 ~75 ms subcortical pulse with cortical confirmation window). CeA mode_prior + fast_prime are injected into SalienceCoordinator via update_signal each select_action tick. BLA retrieval_bias / remap_signal hippocampal consumer wiring is deferred until V3-EXQ-474 confirms behavioural signature. Validation: V3-EXQ-473 CeA mode-prior PASS (5 acceptance criteria), V3-EXQ-474 BLA encoding+remap PASS (5 acceptance criteria), both substrate-readiness diagnostics per EXQ-445 lesson. 33/33 preflight+contract tests PASS with use_amygdala_analog=False (backward compat preserved). Governance: MECH-046 / MECH-074 / MECH-074a/c/d / SD-035 show hold_pending_v3_substrate pending completion. Design doc: docs/architecture/sd_035_amygdala_analog.md; literature synthesis: evidence/literature/targeted_review_amygdala_analog/ synthesis.md.
• MECH-266 asymmetric per-mode hysteresis LANDED 2026-04-21. Schmitt-trigger per-mode enter_thresholds / exit_thresholds dicts layered atop the MECH-259 symmetric switch_threshold in ree_core/cingulate/salience_coordinator.py. Empty-dict default preserves legacy behaviour; over-binding/OCD axis reproducible at exit_threshold near 0 (stuck-in-mode), under-binding/depression axis reproducible with lower enter_threshold. Setters: set_enter_threshold, set_exit_threshold, set_hysteresis_ratio (uniform exit-rail convenience). Validation: V3-EXQ-464 competing-goals 5-arm + V3-EXQ-467 mode-stickiness 5-arm parametric sweep both smoke-PASS all sub-tests. Full behavioural competing-goals runs deferred pending CausalGridWorldV2 dual simultaneously active resource-cue extension.
• SD-029 curriculum-level balanced hazard-event support LANDED 2026-04-21. scheduled_external_hazard_enabled + scheduled_external_hazard_interval + scheduled_external_hazard_prob + scheduled_external_hazard_adjacent_only flags in CausalGridWorldV2 schedule hazard injection (relocate or spawn) at cells adjacent to the agent (or any empty cell when adjacent_only=False). Preserves the self- vs externally-caused taxonomy: the agent did not initiate the encounter. info["external_hazard_injected"] / info["external_hazard_event_count"] tags always present. Unblocks C3/C4 event-conditioned SNR measurement for the MECH-256/SD-029 comparator track which had been failing on per-seed event-count imbalance. Validation: V3-EXQ-470 SCHEDULED vs BASELINE ablation queued.
• SD-033e frontopolar-analog V4-reserved stub LANDED 2026-04-21. ree_core/pfc/frontopolar_analog.py (FrontopolarAnalog + FrontopolarConfig) mirrors the SD-033a contract: no-op behind use_frontopolar_analog=False; raises NotImplementedError when enabled (until design doc lands). Last nn.Linear of both heads (MECH-264 counterfactual-value, MECH-265 relative-importance) zero-initialised. tests/contracts/test_sd_033e_stub.py 7-contract test added (importable, default backward-compat, enabled-raises-NotImplementedError, zero-init, reset safety, get_state stub marker). Three V4-reserved experiment proposals appended to manual_proposals.v1.json (EXP-0165 / 0166 / 0167). Path-clear for the design-doc + dual-active-goal env extension that unlocks behavioural testing.
• Hippocampal anchor-vs-probe cluster REGISTERED 2026-04-21. MECH-269 (regional-verisimilitude anchor selection in hippocampal proposer – per- stream V_s gates anchor eligibility; probe channel inverts the gate for curiosity-driven seeding; anchored rollouts update ARC-018 viability map, probes do not until realized-experience validation); MECH-270 (ephaptic field coherence as candidate biological substrate for V_s readout); MECH-271 (MECH-094 hypothesis_tag as routing signature: anchored replay routes to E1 consolidation / SD-033a PFC, probe replay routes to BLA / NAc – tag is a routing flag, not a source-side marker). MECH-269 lit-pull confidence jumped 0.783 -> 0.852 on 2026-04-21 after Pfeiffer & Foster 2013 (direct evidence: hippocampal sequences start at current location, progress to goal, compositional) was added. Design doc: docs/architecture/hippocampal_anchor_selection.md. MECH-270 future directions: standalone-paper candidate.
• Sleep/waking state-gated routing REGISTERED 2026-04-21. MECH-272 (state-gated anchor/probe routing: waking=anchor-dominant decision-support; sleep=probe-dominant Bayesian schema restructuring); MECH-273 (sleep- dependent aggregation of SD-003 single-episode self-attribution into stable self-model); MECH-274 (V4-reserved: other-attribution sleep-dependent aggregation via ARC-010 empathy / mirror-modelling; implementation_phase: v4). Design doc hippocampal_anchor_selection.md extended with sleep/waking section and V4 other-attribution reservation.
• Scientist-agent developmental-ordering cluster REGISTERED 2026-04-21. ARC-059 (three-stage developmental ordering self -> objects -> others refining ARC-019), MECH-275 (sleep-phase general Bayesian aggregation mechanism – MECH-273/274 become specialisations of MECH-275), MECH-276 (scientist-agent principle: waking-phase counterfactual-backed attribution via deliberate intervention), MECH-277 (action-space discovery via motor experimentation, stage-1 specialisation), MECH-278 (object-schema formation via experimental action, stage-2 specialisation). Design doc: docs/architecture/scientist_agent_developmental_ordering.md. MECH-273/274 depends_on updated with MECH-275.
• Experiment counts: 525 total completions (runner_status.json last_updated 2026-04-22T01:11Z): 108 PASS / 234 FAIL / 62 ERROR / 121 UNKNOWN, up from 495 / 105 PASS / 227 FAIL at the 2026-04-20 snapshot. New PASSes since the 2026-04-21 snapshot: V3-EXQ-473 SD-035 CeA mode-prior, V3-EXQ-474 SD-035 BLA encoding+remap, V3-EXQ-455 SD-032a coordinator behavioural, plus V3-EXQ-456 SD-033a landing (PASS), V3-EXQ-460/466 SD-034 landings, V3-EXQ-462/465 MECH-267, V3-EXQ-463/468 MECH-268, V3-EXQ-464/467 MECH-266. New FAILs since the 2026-04-21 snapshot include V3-EXQ-397c (ARC-007 harder env, 2 attempts), V3-EXQ-445b/c (SD-032b variants), V3-EXQ-133 (MECH-091), V3-EXQ-126 (MECH-104), V3-EXQ-325d (SD-032c AIC), V3-EXQ-452/453/454 (MECH-257 / MECH-261 landing / ARC-016 adaptive), V3-EXQ-433a (MECH-256/ SD-029 comparator scripted-eval). Fresh indexer rebuild pending.
• Pending review count: 46 (as of pending_review.md generation 2026-04-21T19:54:57Z): 8 PASS / 19 FAIL / 19 ERROR-UNKNOWN-smoke. Governance cycle pending to absorb the SD-032 cluster behavioural FAILs, the SD-035 amygdala landings, the MECH-266 hysteresis extension, and the MECH-269/270/271 + MECH-272/273/274 + MECH-275/276/277/278 + ARC-059 registrations.
• Queue drained to 3 items – all claimed. V3-EXQ-447 (SD-032d deterministic validation, ree-cloud-2, claimed 2026-04-19), V3-EXQ-451 (Q-034 hazard/resource threshold retest, EWIN-PC, claimed 2026-04-20), V3-EXQ-445a (SD-032b dACC full-pipeline fix, EWIN-PC, claimed 2026-04-20). All other queued entries from the 2026-04-21 snapshot have since landed as PASS/FAIL entries in runner_status.json.
• Current bottleneck. SD-032 cluster behavioural follow-through remains the primary first-paper-gate blocker: V3-EXQ-445a is the decisive test after the 445b / 445c monostrategy + terrain-inversion variants FAILed. SD-003 successor track (V3-EXQ-433a MECH-256/SD-029 FAIL, V3-EXQ-452 MECH-257 dual-function diagnostic FAIL) is alive. ARC-007 path-memory track remains open (V3-EXQ-397c claimed on DLAPTOP-4.local). SD-035 first-pass hippocampal consumer wiring for BLA retrieval_bias / remap deferred until V3-EXQ-474 behavioural signature confirmed; that work and the MECH-266 full behavioural competing-goals arm both depend on the CausalGridWorldV2 dual simultaneously active resource-cue extension.

Immediate Work Queue (This Cycle)

• Design and queue the EXQ-490b/MECH-295 combined-cluster successor. Post-2026-04-28-governance, EXQ-490b alone is inconclusive: Q-040a substrate-wiring effective PASS at the threshold-overridden smoke; Q-040b stale-stream-discrimination FAIL points at the MECH-295 drive->liking->approach bridge as the remaining blocker. The MECH-295 bridge landed 2026-04-26 (V3-EXQ-493 PASS), so the successor combines MECH-269b VsRolloutGate ON with MECH-295 bridge ON in a single factorial. Resolves the EXQ-483 wired-but-inert pattern by isolating the dominant cause of the observed approach_commit collapse and is the next-up substrate-validation run before V3-EXQ-495 commits.
• Land V3-EXQ-495 (MECH-163 V3 full-completion gate – VTA / hippocampally-planned arm) once the EXQ-490b/MECH-295 successor resolves. All three substrate prerequisites already cleared 2026-04-27 (SD-039 population layer V3-EXQ-494 PASS; MECH-292 V3-EXQ-496 PASS; MECH-293 V3-EXQ-497 PASS). 3 conditions (HABIT / PLANNED / ABLATED) × 2 paradigms (A_DETOUR / B_NOVEL_CONTEXT) × 7 seeds. **C2 PLANNED-HABIT benefit-post-block gap >= 0.30 in detour,

  = 4/7 seeds is THE V3-full-completion criterion.** Estimated ~25h on Mac / ~40h on ree-cloud-1; machine_affinity=any. Queueing-and- running is a deliberate runtime-budget decision.

• Escalate OCD post-Layer-1 disconfirmation to Layer 2 (MECH-290 ablation diagnostic) or Layer 3 (SD-046 multi-slot GoalState pull-forward). V3-EXQ-498 reclassified non_contributory in the 2026-04-28T23:04Z governance cycle (Layer 1 closure-threshold sweep produced no entropy delta vs DEFAULT in 2/3 seeds at any of TIGHT / LOOSE / VERY_LOOSE rails). Layer 2 tests whether ablating MECH-290 backward credit sweep recovers behavioural diversity; Layer 3 tests whether multi-slot GoalState pull-forward resolves the monostrategy.
• Resolve SD-016 env-entropy precondition by extending CausalGridWorldV2 beyond the current SD-023 landmarks-on path. EXQ-418f/g/h established that the cue_context machinery works as designed but the env doesn’t supply cross-context z_world variance (cos~0.998 batch-wise; H1 cos_cross<0.95 not robust at landmarks-on alone). substrate_queue status: parked_pending_env_entropy_precondition. Once env enrichment lands, queue EXQ-418i 4-arm reusing EXQ-418g substrate matrix on the enriched env config.
• Run the next governance cycle once the EXQ-490b/MECH-295 successor + V3-EXQ-495 land: rebuild claim_evidence.v1.json, regenerate pending_review.md, ingest the V3-full-completion-gate outcome. Resolve the open promotion blockers: MECH-294 within-cycle-vs-cross- cycle binding (Kay 2020 challenge); MECH-295 strong-vs-weak liking-bridge necessity (weak reading committed provisionally).
• Pending re-queue under fresh IDs: V3-EXQ-433d / V3-EXQ-418e / V3-EXQ-490 / V3-EXQ-498 successors – only when their predecessor non_contributory / does_not_support reclassifications resolve via substrate progress (Phase 2 forward-predictor V_s for MECH-269b; div_weight sweep at 1.0 / 2.0 / 5.0 for SD-016 path-1; MECH-269/ MECH-269b V_s landing for SD-029 monomodal phenotype; Layer 2 / Layer 3 for OCD post-Layer-1-disconfirmation).
• Aggregator floor flag (5th consecutive cycle): worth a cap-aware aggregator review. Per-paper confidences for low/medium-anchored Q-claims average 0.55-0.86 but claim-level lit_confidence aggregates to 0.82-0.89. Flagged in 6 narrow-open Q-claim evidence_quality_note refreshes this cycle.
• Add a held_v4_by_architectural_commitment recommendation type to the governance tooling so v4-by-design claims (ARC-053 / MECH-225 / MECH-226 / MECH-227 / MECH-228 / MECH-270 / MECH-274 / MECH-276 / MECH-277 / MECH-278 / ARC-059) stop producing the misleading hold_pending_v3_substrate recommendation.
• First-pass hippocampal consumer wiring for SD-035 BLA retrieval_bias and remap_signal once V3-EXQ-474 behavioural signature confirmed (deferred from 2026-04-21 landing pass).
• Continue Sleep Aggregation Cluster wiring: BG / E3 replay-prio consumers reading MECH-284 staleness alongside MECH-269 V_s; downstream consumers of the MECH-272 RoutingGate (HippocampalRouter / E1 ContextMemory) and MECH-273 SelfModelAggregator outputs.
• Move MECH-266 OCD/depression-axis competing-goals behavioural variants (EXQ-464b / EXQ-467b) off hold once the CausalGridWorldV2 dual simultaneously active resource-cue env extension lands.
• SD-033e design doc + dual-active-goal env extension to unlock the three V4-reserved proposals (EXP-0165 / 0166 / 0167).

---

Status Snapshot (2026-04-21)

• SD-033a lateral-PFC-analog / MECH-261 primary consumer LANDED 2026-04-20. ree_core/pfc/lateral_pfc_analog.py (LateralPFCAnalog, LateralPFCConfig). Instantiates MECH-262 rule-selective persistence: gate-modulated EMA rule_state ([1, rule_dim]) with eff_eta = update_eta * write_gate(“sd_033a”), source = delta_proj(z_delta) + world_pool_weight * world_proj(z_world). Frozen-random bias head with last nn.Linear zeroed at init -> initial bias exactly zero (bit-identical with head untrained; training-dependent emergence deferred). Per-mode gate weights from the MECH-261 spec table: external_task=1.0, internal_planning=1.0, internal_replay=0.05, offline_consolidation=0.3. V3-EXQ-456 landing diagnostic PASS (five sub-tests: instantiation, gate-modulated update rate, bias reaches E3 with zero-init contract, backward compat, reset clears rule_state). Design doc: docs/architecture/sd_033a_lateral_pfc_analog.md.
• SD-034 governance closure operator + MECH-268 dACC conflict-saturation + MECH-267 mode-conditioned hippocampal proposals LANDED 2026-04-20..2026-04-21. SD-034 closure operator (ree_core/governance/closure_operator.py) coordinates a five-part “done” token at rule-completion events: (a) MECH-090 beta release, (b) MECH-260 No-Go FIFO injection on the just-completed action class, (c) ResidueField.discharge_domain(z_world, factor, radius) rule-domain multiplicative RBF decay with 1e-6 sign-aware floor (invariant: residue cannot be erased), (d) SalienceCoordinator closure_event signal re-biasing affinity toward internal_planning, (e) dACC PE reset / optional pe_cap install (MECH-268). Completion detector: rule_state delta < threshold for N consecutive ticks AND beta elevated AND current_mode in allowed_closure_modes AND write_gate(“sd_033a”) >= min. Mode conditioning is the falsifiability predicate vs pure MECH-090 + MECH-260 + MECH-094 tuning. MECH-267 (ree_core/hippocampal/module.py) threads operating_mode through HippocampalModule.propose_trajectories with per-mode CEM-noise multipliers. MECH-268 (ree_core/cingulate/dacc.py) adds an outcome-history FIFO + f_sat attenuation on the dACC bundle; closure_event resets the buffer. Landing smokes all PASS: V3-EXQ-460 SD-034 closure wiring (6 sub-tests), V3-EXQ-466 ResidueField.discharge_domain (5 sub-tests: near attenuation, far spared, invariant preserved, end-to-end, distant-z spares), V3-EXQ-462 MECH-267 rule binding, V3-EXQ-465 MECH-267 intrusive-simulation filtering, V3-EXQ-463 + V3-EXQ-468 MECH-268 saturation-and-reset. Behavioural variants with full E3 task loop + tolerance-band completion env deferred (depend on phased rule_state training + env variant not yet on the roadmap). Anchor: evidence/planning/sd033_governance_plan.md; source: docs/thoughts/2026-04-20_ocd4.md + GAP MEMO “REE-V3 is not missing cognition, it is missing governance.”
• SD-032 cluster behavioural follow-through: FAIL across four first-pass behavioural gates. V3-EXQ-445 FAIL (SD-032b 3-arm ablation hit the monostrategy + terrain-inversion fishtank_viz signature under all three configs; dACC score_bias entropy delta under the C2 gate); V3-EXQ-325d FAIL (SD-032c AIC descending modulation, does_not_support); V3-EXQ-454 FAIL (ARC-016 adaptive commitment threshold, weakens). V3-EXQ-455 PASS (SD-032a salience-network coordinator behavioural: supports SD-032a / MECH-259 / MECH-261 on the synthetic high-PE injection path). V3-EXQ-452 FAIL (MECH-257 dual-function E2 diagnostic), V3-EXQ-453 FAIL (MECH-261 write-gate landing diagnostic – SD-032e-relevant). Net reading: the salience-coordinator substrate and its write-gate registry are structurally sound in isolation, but the end-to-end dACC / AIC / ARC-016 behavioural loop does not yet clear even the first behavioural gates on CausalGridWorldV2. EXQ-445 has three successors queued (a/b/c) targeting monostrategy + terrain inversion via MECH-260 suppression, ARC-058 shared-trunk, and foraging-value wiring respectively. EXQ-325b re-scoped as EXQ-325d produced does_not_support; AIC->descending pathway remains open under drive-regime contrast.
• SD-016 forward-path diagnostic: V3-EXQ-449 FAIL confirmed cue_action_proj receives exactly 0.0 gradient under the original “implicit via E3 trajectory selection” claim (C1 PASS, 2 seeds, ~1.7k steps; CEM argmax non-differentiable + detach at agent.py:694). C2 arm added supervised MSE loss against E2.action_object(z_world, a_executed) .detach() – weights trained (grad ~0.013, delta ~0.21) but action_bias_divergence stayed at 0.0, indicating a downstream blocker between cue_action_proj and E3.select. V3-EXQ-449a queued to instrument the full forward path and identify the specific blocker before any EXQ-418b successor is written. cue_action_proj is now treated as CURRENTLY UNGROUNDED: sd016_enabled=True experiments should expect action_bias_divergence ~= 0.0 on the action path; cue_terrain_proj remains valid (trained via terrain_loss).
• Recent landing MECH-267 + MECH-268 substrate smokes all PASS. V3-EXQ-462 (MECH-267 rule binding) supports [MECH-267, SD-033a, MECH-262]. V3-EXQ-465 (MECH-267 intrusive-simulation filtering) supports [MECH-267, MECH-094, MECH-261]. V3-EXQ-463 + V3-EXQ-468 (MECH-268 conflict-saturation) supports outcome-history FIFO + f_sat attenuation + closure-event buffer reset. V3-EXQ-456 (SD-033a landing) supports [SD-033a, MECH-261, MECH-262].
• ~715 V3 runs indexed (indexer rebuild 2026-04-20T19:49Z wrote claim_evidence.v1.json with 630 V3 run-dirs + 77 flat V3 manifests = 707 post-epoch; ~9 further V3 manifests written since rebuild covering SD-033a / SD-034 / MECH-267 / MECH-268 landings and the SD-032 cluster behavioural follow-through). Fresh claim_evidence.v1.json rebuild pending after this cycle’s wave of results. Queue at snapshot time: 14 active items, 4 claimed – V3-EXQ-447 (SD-032d ree-cloud-2), V3-EXQ-451 (Q-034 retest EWIN-PC), V3-EXQ-445a (SD-032b full-pipeline fix EWIN-PC), V3-EXQ-397c (ARC-007 harder-env DLAPTOP-4). Pending queue: V3-EXQ-445b/c (SD-032b variants), V3-EXQ-456 (SD-033a landing, now PASS), V3-EXQ-449a (SD-016 forward-path probe), V3-EXQ-133 / V3-EXQ-126 (MECH-091 / MECH-104 discriminative pairs), V3-EXQ-460 / 463 / 466 / 468 (SD-034 + MECH-268 landing smokes, all PASS). runner_status.json last_updated stale at 2026-04-20T14:39:30Z (495 completions: 105 PASS / 227 FAIL / 62 ERROR / 101 UNKNOWN); the live machine-side completion log is ahead of that snapshot.
• Pending review count: 6 (as of pending_review.md generation at 2026-04-20T05:50:27Z; stale – regeneration pending after this cycle). Items: FAIL EXQ-397 (ARC-007/SD-004 path memory), FAIL EXQ-433a (MECH-256/SD-029 scripted-eval comparator), FAIL EXQ-445 (SD-032b behavioural); PASS EXQ-446 (SD-032a coordinator landing); ERROR V3-EXQ-445 + V3-EXQ-325c to clear.
• Governance cycle 2026-04-19T21 (post-SD-032 landing) carry-forward. Promoted MECH-094 to provisional; applied 12 hold_pending_v3_substrate decisions for the SD-032 cluster and dependents; reclassified EXQ-395 / EXQ-418a / EXQ-430 as non_contributory substrate-gap symptoms. No new governance cycle run today – the SD-032 behavioural FAILs and the SD-033a/SD-034/MECH-267/MECH-268 landings are the input set for the next cycle.
• Current bottleneck: SD-032 cluster behavioural escape from monostrategy + terrain-inversion, SD-033 governance cluster behavioural validation, SD-016 forward-path blocker identification, SD-003 successor track. Regression suite PRs 1-5 landed (preflight + contracts + deferred changed; /api/regression/preflight serve.py endpoint; explorer preflight badge; pre-commit contracts hook).

Immediate Work Queue (This Cycle)

• Land results for the four claimed experiments: V3-EXQ-447 (SD-032d deterministic validation, ree-cloud-2), V3-EXQ-451 (Q-034 retest, EWIN-PC), V3-EXQ-445a (SD-032b full-pipeline fix, EWIN-PC), V3-EXQ-397c (ARC-007 path memory harder-env, DLAPTOP-4).
• Review pending_review.md after its next regeneration – expected to cover EXQ-397 / EXQ-433a / EXQ-445 FAILs plus EXQ-446 PASS and the two ERROR clears (V3-EXQ-445, V3-EXQ-325c).
• Queue and land V3-EXQ-449a (SD-016 forward-path instrumentation probe) as the prerequisite for any EXQ-418b successor.
• Land V3-EXQ-445b/c (SD-032b monostrategy + terrain-inversion variants) once V3-EXQ-445a returns.
• Behavioural variants for SD-034 / MECH-267 / MECH-268 still need a tolerance-band completion env + phased rule_state training plan before any behavioural EXQ can be written; the landing-diagnostic smokes (V3-EXQ-460/462/463/465/466/468) have all PASSed.
• Next governance cycle: ingest the SD-032 behavioural FAILs, the SD-033 cluster landings, and the MECH-094 provisional persistence; rebuild claim_evidence.v1.json; regenerate pending_review.md.

---

Status Snapshot (2026-04-20)

• SD-032 cingulate integration cluster fully IMPLEMENTED 2026-04-19 (a/b/c/d/e). In order of landing: SD-032b dACC/aMCC-analog adaptive control (Croxson/Shenhav/Kolling bundle -> DACCtoE3Adapter shim -> E3.select score_bias; ARC-033 vs ARC-058 shared-trunk as constructor-switch alternative), then SD-032a salience-network coordinator (soft operating_mode vector + MECH-259 Schmitt-trigger switch threshold + MECH-261 dict-keyed write-gate registry, 8 default targets, V4 register_target() extensibility), then SD-032c AIC-analog (drive- and mode-aware harm_s_gain subsumes SD-021 descending modulation; EXQ-325a bit-identical DESCENDING==CONTROL signature resolved), then SD-032d PCC-analog metastability scalar (modulates MECH-259 effective_threshold by drive_level / success EMA / time-since-offline; single integration point for MECH-092 within-session quiescence and INV-049 cross-session sleep via enter_offline_mode), then SD-032e pACC-analog slow-EMA autonomic coupling (drive_bias write-back from z_harm_a, MECH-094 hypothesis_tag gated, alpha=0.002 default ~347-step half-life inside Guo 2018 ACC mGluR5 LTP envelope). All modules under ree_core/cingulate/, backward-compatible master switches default False.
• MECH-094 promoted candidate -> provisional (governance-2026-04-19T21). First concrete write-gate wiring established by V3-EXQ-448 pACC hypothesis_tag skip PASS; 12 supports / 0 opposing, confidence 0.856. Feeds the MECH-261 mode-conditioned generalisation.
• SD-033 PFC subdivision cluster registered 2026-04-19. SD-033 parent + SD-033a-e (lateral-PFC / premotor-analog / vmPFC-analog / OFC-analog / frontopolar parallel-goal deliberation) + MECH-262/263 + MECH-264/265 (frontopolar counterfactual-value and relative-importance monitoring). V3-pending; primary write target for MECH-261 operating-mode-conditioned writes. Prong D frontopolar lit-pull (6 entries, mean conf 0.81, Boorman 2009 / Mansouri 2017 load-bearing) broadened SD-033e from Koechlin branching to parallel-goal deliberation; reserved V4 operating-mode renamed deliberative_branching -> parallel_goal_deliberation (zero schema cost — mode names are dict keys). Design docs: docs/architecture/sd_032_cingulate_integration_substrate.md, docs/architecture/sd_033_pfc_subdivision_architecture.md.
• Regression suite PRs 1-3 landed (ree-v3). Three-layer architecture: (1) preflight (tests/preflight/, runner imports + queue integrity + machine boot; wired into experiment_runner.py startup with --skip-preflight escape hatch); (2) contracts (tests/contracts/ with C1 agent boot, C2 8-flag boot matrix, C3 seed determinism, C4 BG gating MECH-090/091, C5 imagined/acted isolation MECH-094, C6/C7/C8 SD-032 dACC/AIC/PCC/pACC wiring; 24/24 pass in ~14s); (3) deferred changed layer stubbed via scripts/run_regression_suite.py. Serve.py /api/regression/preflight endpoint with 60s memoisation added (REE_assembly commit 2cb1c9559). Contracts test wiring only, never thresholds from EXQ evidence.
• EXQ-433 reclassified non_contributory; EXQ-433a scripted-eval successor queued (2026-04-19). Root cause of EXQ-433 FAIL: event-distribution collapse in 3/4 seeds (seed 91: 303/0 agent/env, seeds 13/42: ~100/1-2). MECH-256 C1 forward_r2=0.983-0.9998 unaffected. EXQ-433a uses CausalGridWorldV2.reset_to() for deterministic placement + 30-trial scripted harness per event type; balanced 3/3/3/3 event counts in smoke; supersedes EXQ-433.
• Governance cycle 2026-04-19T21 (post-SD-032 landing). Promoted MECH-094 to provisional; held SD-020 at provisional; applied 12 hold_pending_v3_substrate batch (MECH-256/258/259/260/261/264/265, SD-029/032a/b/d/e); reclassified 3 FAIL manifests as non_contributory substrate-gap (EXQ-395 MECH-220, EXQ-418a SD-017, EXQ-430 INV-010 — all addressable by SD-032 cluster); marked 7 experiments reviewed. Retest eligibility post-cluster: V3-EXQ-325b (SD-032c falsification signature); V3-EXQ-430a (MECH-261 offline-write-gating); V3-EXQ-418b pending diagnostic of SD-016 cue_action_proj wiring. 5 SD-032 entries added to substrate_queue.json as implemented; EXP-0121 and EXP-0132 marked executed.
• MECH-261 mode-gating lit-pull (2026-04-20T06:30Z). 5 entries (Latchoumane 2017 SO-spindle-ripple triple coupling; Maingret 2016 hippocampo-cortical coupling reorganises mPFC; Helfrich 2018 MFC atrophy disperses SO-spindle coupling; Klinzing/Niethard/Born 2019 review; Boyce 2016 REM theta optogenetic). V4-staging findings: per-mode gate weights load-bearing; carrier rhythm is biological realisation of gate in SWS and REM; gate locus and write target may overlap in mPFC; within-REM target selection remains open. 795 literature entries, 805 runs indexed at lit-pull rebuild.
• 704 runs indexed (630 dirs + 74 flat); 831 queue-level completions across five machines (252 PASS / 480 FAIL / 88 ERROR / 11 UNKNOWN). Queue: 2 items, both claimed — V3-EXQ-445 (SD-032b 3-arm ablation OFF / ON-independent / ON-shared-trunk, DLAPTOP-4.local) and V3-EXQ-447 (SD-032d deterministic validation, ree-cloud-2).
• Current bottleneck: SD-032 cluster behavioural validation + SD-003 successor follow-through. V3-EXQ-445 is the first behavioural gate on SD-032b (C2 dACC score_bias produces >=0.1 nats entropy delta in either ON arm). V3-EXQ-433a supersedes EXQ-433 as the MECH-256/SD-029 comparator test on scripted balanced events. 3 pending review (EXQ-397 FAIL ARC-007/SD-004 path memory, EXQ-433a FAIL MECH-256/SD-029 scripted-eval comparator, V3-EXQ-325c ERROR).

Immediate Work Queue (This Cycle)

• Land results for V3-EXQ-445 (SD-032b behavioural) and V3-EXQ-447 (SD-032d deterministic); both are the first post-landing validation gates for the cingulate cluster.
• Review EXQ-397 (ARC-007/SD-004 path memory) and EXQ-433a (MECH-256/SD-029 scripted comparator) in the next governance pass; clear V3-EXQ-325c ERROR.
• Queue SD-032 cluster retests now unblocked by substrate arrival: V3-EXQ-325b (SD-032c falsification signature), V3-EXQ-430a (MECH-261 offline-write-gating), and V3-EXQ-418b (gated on SD-016 cue_action_proj wiring diagnostic).
• MECH-261 V4 staging decisions tracked against the 2026-04-20 mode-gating lit-pull (carrier-rhythm gate implementation, within-REM target selection).

---

Status Snapshot (2026-04-18)

• SD-003 superseded. After 28 accumulated FAILs across the two-pass counterfactual architecture, SD-003 was flipped to superseded with superseded_by: [MECH-256, SD-029]. New claims registered: MECH-256 (general single-pass forward-model comparator, stream-agnostic; Frith/Shergill/Haggard/Blakemore biology), MECH-257 (dual-function single-substrate E2: comparator vs evaluator, controller-gated), SD-029 (concrete z_harm_s instantiation of MECH-256), SD-030 (z_self stream, V4-deferred), SD-031 (z_world stream, V4-deferred). Architecture doc: docs/architecture/self_attribution_per_stream.md. claims.yaml: 491 claims (+5).
• V3-EXQ-433 queued (SD-003 successor test, next-up). Event-conditioned single-pass comparator test on z_harm_s: residual = z_harm_s_observed − E2harm_s(z_harm_s{t−1}, a_actual). SD-013 interventional training (fraction=0.5) during P1; P2 uses event-density controller that extends up to 200 episodes until ≥20 env-caused and ≥20 agent-caused hazards per seed (fixes EXQ-431 sample starvation). Criteria: C1 forward_r2 ≥ 0.9, C2 self/ext attenuation ratio ∈ [0.3, 0.7] (Shergill), C3 approach SNR > 3, C4 density floor; PASS needs 3/4 seeds. Substrate prerequisites (ARC-033, SD-013, SD-011) all implemented — SD-029 is a read-mode claim over existing substrate.
• Governance cycle 2026-04-18 (governance-2026-04-18-15z). 2 pending_user recommendations applied as hold_pending_v3_substrate: SD-014 (implementation_phase=v3, 4 supports/0 weakens lit-only) and SD-023 (override of indexer’s promote_to_provisional to hold, with EXQ-332a indexed non_contributory). Pipeline clean (validator OK 68/68, 772 runs indexed). Pending review cleared: 0.
• New lit-pulls (wave 1): LIT-0092 (MECH-104 LC-NE volatility; Sara 2009 Nat Rev Neurosci filled triangulation), LIT-0097 (INV-053 depression attractor; Huys/Daw/Dayan 2015 added as HDD contrast class — HDD and INV-053 are complementary not identical). Plus three lit-pulls informing SD-003 successor design (comparator, evaluator, mode distinction modes). Literature entries: 741 (+14).
• New experiments queued (wave 2): V3-EXQ-434 (INV-053 depression attractor replication, 5-seed LONG_HORIZON), V3-EXQ-435 (INV-054 phase-transition recovery, sustained-crossing criterion, supersedes EXQ-278), V3-EXQ-436 (SD-017 sleep phase ablation redesign with context-conditioned harm threshold, supersedes EXQ-242).
• ree-cloud-2 onboarded. Second Hetzner cloud worker (CX22 nbg1, IPv4 116.203.216.181) brought online. Parameterised systemd service template; cloud-scaler.yml extended to a two-server loop (ree-worker-1/ree-cloud-1 + ree-worker-2/ree-cloud-2); validator whitelist extended; contributor JSON registered. First real claim was V3-EXQ-355b (ARC-038 schema assimilation) rather than the dedicated smoke, because the runner’s iteration order put the smoke behind any-affinity items — de facto pipeline verification.
• 772 runs indexed; 517 queue-level completions in runner_status.json (102 PASS / 238 FAIL / 63 ERROR / 114 UNKNOWN).
• Current bottleneck: SD-003 successor architecture validation + first-paper gate. Active queue (17 items): V3-ONBOARD-smoke-ree-cloud-2, V3-EXQ-433 (SD-029 event-conditioned comparator, next-up), V3-EXQ-326, V3-EXQ-330a, V3-EXQ-328b, V3-EXQ-326a, V3-EXQ-407, V3-EXQ-332, V3-EXQ-321c, V3-EXQ-325b, V3-EXQ-355b, V3-EXQ-418b, V3-EXQ-434, V3-EXQ-435, V3-EXQ-436, V3-EXQ-406b, V3-EXQ-429b.

---

Status Snapshot (2026-04-17)

• New substrate: SD-016 (frontal cue-indexed integration) implemented 2026-04-16. E1 queries ContextMemory via z_world using world_query_proj; cue_action_proj provides affordance bias to E2; cue_terrain_proj provides (w_harm, w_goal) terrain precision weights to E3. Config: E1Config.sd016_enabled (default False, backward compatible). Design doc: REE_assembly/docs/architecture/sd_016_frontal_cue_integration.md. Validation experiment V3-EXQ-418a queued with terrain_loss fix.
• Governance 2026-04-16 completed: 16 experiments reviewed. 5 PASS: EXQ-049a (MECH-090 bistable concordance, Layer 1+2 regression), EXQ-365 (MECH-104 surprise gate, 5-seed), EXQ-353 (ARC-033/SD-003/SD-013 interventional vs observational counterfactual), EXQ-323a (SD-019 affective nonredundancy on SD-022 substrate), EXQ-328a (MECH-090 bistable + SD-012). 11 FAIL/non_contributory/inconclusive including EXQ-385/418 (INV-049/SD-017, identical per-seed data — SHY collapse root cause identified), EXQ-355 (ARC-038 optimizer contamination), EXQ-330a (SD-013, later PASS in EXQ-330a — already confirmed 2026-04-15), EXQ-324a (SD-020 inconclusive, eval termination bug). 9 manifest reclassifications. 4 fix scripts written and queued (EXQ-418a, EXQ-385a, EXQ-355a, EXQ-324b).
• EXQ-321a FAIL (2026-04-17): MECH-090 bistable gate still failing. Root causes: 4-bug chain (shared training deepcopy on autograd tensors, spike timing vs E3-tick alignment, bistable config silently dropped via **kwargs). EXQ-321b queued with all 4 fixes; dry-run 3/3 seeds PASS.
• 766 runs indexed (per morning digest 2026-04-17). 2 pending review (EXQ-321a FAIL + UNKNOWN runner entry).
• SD-016 lit-pull (2026-04-17): Additional 3 entries added to targeted_review_sd_019 (wind-up/central sensitization, Craig 2003 interoception/insula, pain asymbolia). SD-022 lit-pull added 2 entries. Index now 706 lit entries total.
• Deployment gating note added (2026-04-17): V3 is treated as a sandbox-only scientific substrate. High-capability or externally connected REE deployment is gated on V4 social/developmental completion. See docs/governance/deployment_gating.md.
• Current bottleneck: first-paper gate. Active queue (18 items): EXQ-326 (SD-015/MECH-216/SD-012 wanting nav fix), EXQ-330a (SD-013 claimed), EXQ-321b (MECH-090 bistable fix), EXQ-325a (SD-021 descending modulation claimed), EXQ-395 (MECH-220), EXQ-375 (MECH-073), EXQ-328b (MECH-230 claimed), EXQ-326a (SD-015/MECH-229 nav), EXQ-406 (INV-053), EXQ-407 (MECH-231), EXQ-396 (ARC-016), EXQ-397 (ARC-007), EXQ-429 (INV-044), EXQ-430 (INV-010), EXQ-418a (SD-016+SD-017 fix), EXQ-385a (INV-049 SHY fix), EXQ-355a (ARC-038 optimizer fix), EXQ-324b (SD-020 eval fix).

---

Status Snapshot (2026-04-15)

• New substrate today: MECH-090 Layer 1 (trajectory stepping) + MECH-091 Layer 2 (urgency interrupt) implemented. REEAgent now steps through committed_trajectory.actions[idx] via _committed_step_idx counter (Layer 1). Layer 2: when beta elevated and z_harm_a.norm() > urgency_interrupt_threshold (default 0.8), gate releases and step counter resets. Both wired in agent.py + E3Config.urgency_interrupt_threshold in config.py (2026-04-15).
• New claims registered 2026-04-14: MECH-232 (DA representational expansion as approach mechanism), MECH-233 (asymmetric valence encoding: BLA tags vs VTA expands), ARC-057 (curiosity-approach emergence from DA-expanded map). Architecture doc: hippocampal_valence_asymmetry.md. MECH-231 promoted candidate->provisional (conf from EXQ-407 PASS, 164x E2/E1 slope ratio).
• EXQ-330a PASS (2026-04-15): SD-013 contrastive counterfactual at interventional_fraction=0.5. forward_r2=0.999, cf_gap confirmed. Advances SD-013 evidence (already provisional conf=0.788).
• EXQ-327 PASS (2026-04-14): MECH-163 goal-conditioned navigation paper gate confirmed.
• EXQ-365 PASS (2026-04-14): MECH-104 surprise gate (5-seed) confirmed.
• 494 experiments completed. 100 PASS, 236 FAIL, 51 ERROR, 107 UNKNOWN.
• 0 pending review (as of 2026-04-15).
• Current bottleneck: first-paper gate. Active queue (16 items): EXQ-323a (SD-019 nonredundancy), EXQ-326 (SD-015 wanting gradient nav), EXQ-330a (claimed), EXQ-353 (SD-003 interventional vs observational), EXQ-321a (MECH-090 bistable gate retest), EXQ-325a (SD-021 descending modulation retest), EXQ-395 (MECH-220), EXQ-375 (MECH-073), EXQ-328b (claimed), EXQ-326a, EXQ-406 (INV-053), EXQ-407 (MECH-231), EXQ-396a (ARC-016 dual-bug fix), EXQ-396 (ARC-016 sweep), EXQ-397 (ARC-007 path memory), EXQ-418 (SD-017 + SD-016 context action).

---

Status Snapshot (2026-04-14)

• Key governance outcome: SD-013 promoted candidate->provisional (2026-04-13b governance). conf=0.788, 5 supports/1 weakens. SD-013 (interventional training bias) now provisional. 7 experiments reclassified non_contributory.
• New claim class registered: EXT-001 through EXT-007. External AI/LLM failure mode catalogue with REE mechanism mappings (sycophancy, hallucination, reward hacking, goal misgeneralization, causal attribution gap, other-model collapse, context amnesia). claims.yaml: 454 claims total.
• MECH-231 registered (E2 short-horizon discriminative pair, cowork-2026-04-13-e). EXQ-407 queued.
• EXQ-406 queued (INV-053 depression attractor replication, 5-seed LONG_HORIZON characterisation, ~240 min).
• ~481 experiments completed. 96 PASS, 235 FAIL, 51 ERROR, 99 UNKNOWN.
• 2 pending review (as of 2026-04-14T04:18:29Z): v3_exq_326_wanting_gradient_nav_fix FAIL (MECH-216/SD-012/SD-015); V3-EXQ-326 UNKNOWN.
• Current bottleneck: first-paper gate. Active queue (17 items): EXQ-326, EXQ-332, EXQ-330a, EXQ-353, EXQ-322a, EXQ-328a, EXQ-321a, EXQ-325a, EXQ-365, EXQ-355, EXQ-395, EXQ-375, EXQ-385, EXQ-328b, EXQ-326a, EXQ-406 (INV-053 depression attractor), EXQ-407 (MECH-231 E2 short-horizon).

---

Status Snapshot (2026-04-13)

• Key governance outcome: EXQ-354 PASS (2026-04-13). MECH-112 behavioral wanting/liking dissociation confirmed with SD-015 wiring (3/3 seeds). MECH-112 split into MECH-229 (behavioral wanting/liking dissociation, active) and MECH-230 (latent z_goal structure, candidate). 7 dry-run FAILs from earlier sessions reclassified as non_contributory. SD-012 design doc updated to IMPLEMENTED.
• Five new experiments queued (2026-04-13): V3-EXQ-355 (ARC-038 schema assimilation), V3-EXQ-365 (MECH-104 surprise gate, 5-seed), V3-EXQ-375 (MECH-073 valence geometry), V3-EXQ-385 (INV-049 offline consolidation necessity / sleep ablation), V3-EXQ-395 (MECH-220 harm hub behavioral probe). Plus EXQ-328b (MECH-230 full run) and EXQ-326a (SD-015 + MECH-229 nav fix).
• SDs moved pending->implemented since last snapshot: SD-013 (interventional training), SD-015 (ResourceEncoder), SD-017 (minimal sleep infrastructure), SD-018 (resource proximity supervision), SD-019 (affective nonredundancy constraint), SD-020 (affective harm surprise PE), SD-021 (descending pain modulation), SD-022 (directional limb damage), SD-023 (environmental gradient texture). Also: ARC-033, MECH-090 bistable gate, MECH-120 SHY wiring, MECH-203/204 serotonin substrate, MECH-205 surprise-gated replay fix, MECH-216 E1 predictive wanting.
• 481 experiments completed. 96 PASS, 235 FAIL, 51 ERROR, 99 UNKNOWN.
• 0 pending review (as of 2026-04-13T07:19:18Z).
• Current bottleneck: first-paper gate. Active queue (16 items): EXQ-327 (MECH-163 goal-conditioned nav paper gate), EXQ-326a (SD-015 + MECH-229 nav fix), EXQ-353 (ARC-033/ SD-003/SD-013 interventional vs observational counterfactual), EXQ-321a (MECH-090 bistable gate retest), EXQ-325a (SD-021 descending modulation retest), EXQ-365 (MECH-104 surprise gate), EXQ-355 (ARC-038), EXQ-395 (MECH-220), EXQ-375 (MECH-073), EXQ-385 (INV-049 sleep ablation), and fix iterations EXQ-322a/328a/330a/332/328b/326a.

---

Status Snapshot (2026-04-06)

• SD-011/SD-012 Full E3 Integration (2026-04-05). z_harm_a now flows through the complete agent loop: agent.sense() -> LatentStack.encode() -> E3.select(). New E3Config parameters: urgency_weight (z_harm_a.norm() lowers commit threshold, D2 avoidance escape, capped by urgency_max=0.5) and affective_harm_scale (amplifies lambda_ethical by accumulated threat). E3.compute_harm_forward_cost() replaces deprecated HarmBridge path, rolling z_harm_s step-by-step through trajectory actions via ResidualHarmForward. Agent.compute_drive_level(obs_body) added as canonical SD-012 static method. All new parameters default to 0.0/None for full backward compatibility.
• EXQ-247 queued: 4-arm ablation validating full SD-011/SD-012 E3 integration. Tests urgency_weight and affective_harm_scale jointly with drive_weight across ablation conditions (FULL vs NO_URGENCY vs NO_AFFECT vs BASELINE). 3 seeds x 200 train + 50 eval x 200 steps.
• New claims registered (2026-04-06 thought-intake sessions): INV-049 (Offline Update Necessity Principle – offline phases are a mathematical necessity for model-building agents), INV-050 (three-drive sleep regulation), INV-051 (optimal novelty range), MECH-178 (noradrenergic REM suppression pathway), MECH-179 (MEL type-specificity), MECH-180 (novelty-driven adaptive sleep), MECH-181 (cognitive reserve as update-loop maintenance), Q-033 (actigraphy MEL forecasting).
• ~198 experiments run. 51 PASS, 123 FAIL, 22 ERROR, 2 UNKNOWN. 22 experiments currently queued (EXQ-223 through EXQ-247 series plus EWIN-PC onboarding smoke).
• 0 pending review (as of 2026-04-04T18:45:00Z).
• Current bottleneck: first-paper gate. Active queue: EXQ-074e/234 (wanting/liking), EXQ-076e/235 (goal conditioning), EXQ-195 (SD-003 z_harm_s counterfactual), EXQ-247 (SD-011/012 full integration), and sleep-architecture experiments (EXQ-242–246).

---

Status Snapshot (2026-04-04)

• SD-014 implemented (2026-04-04): hippocampal valence vector node recording. 4-component valence vector V=[wanting, liking, harm_discriminative, surprise] added to RBFLayer and ResidueField (ree_core/residue/field.py). Each RBF center stores a valence_vecs buffer [num_centers, 4] updated incrementally per visit. MECH-094 gate applies: hypothesis_tag=True blocks valence updates. Prerequisite for ARC-036 and replay prioritisation via drive state.
• ARC-028 + MECH-105 implemented (2026-04-04): hippocampal-BetaGate completion coupling. HippocampalModule.compute_completion_signal() maps best trajectory score to a sigmoid dopamine-analog value. BetaGate.receive_hippocampal_completion() releases beta when signal

  = threshold (0.75). Implements the Lisman & Grace 2005 subiculum->NAc->VP->VTA loop.

• ~292 experiment scripts authored. EXQ-001 through EXQ-223 series. 0 pending review as of 2026-04-03 (all discussed). EXQ-125 currently running on DLAPTOP-4.local (ARC-029).
• Governance clean. 0 pending review items (generated 2026-04-03T21:39:23Z). Last governance cycle: 2026-04-03 (cowork-2026-04-03-b, 14 experiments reviewed; ARC-022 promoted to provisional).
• Current bottleneck: first-paper gate experiments. EXQ-223 PASS confirmed the minimal E1+E2+hippocampus core loop. Active queue: EXQ-074e (MECH-112/117 wanting/liking), EXQ-076e (MECH-116 E1 goal conditioning), EXQ-195 (SD-003 z_harm_s counterfactual), EXQ-125 (ARC-029 committed mode, running). SD-015 resource indicator in progress.

---

Status Snapshot (2026-04-03)

• EXQ-223 PASS: Minimal mind confirmed (2026-04-03). The REE core loop — E1 (associative world model) + E2 (fast transition model) + HippocampalModule (trajectory proposal) + multinomial go/no-go + raw harm/reward signals — is sufficient for stable navigation, harm avoidance, and resource acquisition. 3/3 criteria met across all 3 seeds (harm_ratio 0.29–0.39; REE takes ~4.5× as much reward as random). The ablation strips the deliberative architecture entirely: commitment_threshold=−1.0 (always uncommitted), z_goal disabled, benefit_eval disabled. What remains is the predictive associative core alone — and it works. This is the first experimental confirmation that the E1+E2+hippocampus triad constitutes a minimal functional mind. The circuit topology is a named-structure match to the zebrafish larva (5–7 dpf): dorsal pallium (E1) → cerebellum (E2) → lateral pallium (hippocampal module) → optic tectum + reticulospinal neurons (go/no-go) → lateral habenula (harm signal). The larva has no mature prefrontal cortex — no commitment architecture — matching the ablation exactly. It is the only vertebrate whose entire ~100,000-neuron CNS has been functionally imaged during free behaviour (Ahrens et al., 2013, Nature Methods; Portugues et al., 2014, Neuron). This match was derived from functional-architecture arguments, not from biology. Episode visualiser (episode_viewer.html) added to the explorer for trajectory playback. Full circuit table and references: see changelog 2026-04-03.
• SD-011 and SD-012 both implemented. SD-011 (dual nociceptive streams: z_harm_s + z_harm_a) validated at EXQ-178b PASS (2026-03-30). SD-012 (homeostatic drive modulation) implemented 2026-04-02: drive_weight default raised from 0.0 to 2.0, enabling effective_benefit = benefit_exposure * (1.0 + drive_weight * drive_level). Step 3.1 substrate debt now substantially resolved (SD-008/009/010/011/012 all done).
• ~198 experiments run. EXQ-001 through EXQ-212+ series. 51 PASS, 123 FAIL, 22 ERROR as of 2026-04-03. Breadth of FAIL reflects aggressive experimentation on a developing substrate – each FAIL cluster was analyzed and resolved before the next iteration.
• Breath oscillator and z_beta pathway wired (2026-04-02). BreathOscillator integrated into core commitment decision (MECH-108). rv -> z_beta volatility pathway wired for Q-007. 14 substrate-limited experiments marked scoring_excluded. EXQ-199–203 queued to re-run MECH-025/Q-007/MECH-029/MECH-026/MECH-057a on corrected substrate.
• ARC-033 ResidualHarmForward promoted to ree_core (2026-04-02). E2_harm_s forward model now in ree_core/latent/stack.py, enabling EXQ-195 (SD-003 z_harm_s counterfactual). EXQ-195 queued; this is the critical SD-003 re-validation step on the new harm pipeline.
• Governance cycle active. 36 experiments pending review (2026-04-03). PASS cluster includes SD-011 validation (EXQ-178b), terrain work (SD-015), and several MECH-1xx claims; FAIL cluster being classified (evidence vs diagnostic). Governance session in progress.
• Current bottleneck: first-paper gate experiments. SD-011/012 substrates cleared; next priority is EXQ-074e (MECH-112/117 wanting/liking), EXQ-076e (MECH-116 E1 goal conditioning), EXQ-195 (SD-003 z_harm_s counterfactual), and EXQ-182a (oracle ceiling for habit-system goal lift).

---

Status Snapshot (2026-03-31)

• V3 first-paper gate clarified. ree-v3 completion for the first paper is now explicitly scoped to the waking, single-agent substrate. Sleep, social extension, language/communication, nth-order ethics, and full psychiatric modelling are not blockers for V3 completion.
• Immediate focus is the approach/goal side. Harm/attribution substrate has advanced materially (SD-003 architecture, SD-005, SD-010, SD-011, ARC-033), but the main remaining V3 risk is still positive attractor behavior rather than harm avoidance.
• Hard V3 completion gates for paper 1: (1) post-SD-011 SD-003 works on z_harm_s, not only the legacy z_world form; (2) harm/attribution substrate is stable enough to treat as platform rather than ongoing rescue work; (3) SD-012 + MECH-112 yield genuine behavioral goal lift, not only z_goal activation; (4) ARC-030 is demonstrated as dual evaluation of the same trajectories by harm and goal channels inside one selector; (5) matched-seed reruns and at least one task variant confirm robustness; (6) governance state remains clean enough that review/index/claim status are aligned.
• First-paper claim remains narrow. Target claim is: REE architectural separation yields attributable, harm-avoiding, goal-directed agency in a waking single-agent substrate.
• Deferred to V4/V5: consolidation/sleep mechanisms, integrated self/other modelling, structured communication, and emergent ethical behavior in multi-agent settings. (Note 2026-04-02: The V4 deferral of social work reflects a precise architectural constraint: INV-043 establishes that testing whether ethical capacity is developmentally activated — not merely architecturally present — requires a multi-agent substrate with modelled caregiving. V3 tests the machinery (ARC-043 Layer 6); INV-043 testing requires Layers 2-4 to be exercised socially. See docs/architecture/developmental_curriculum.md#inv-043.)

---

Status Snapshot (2026-03-26)

• V2 complete. Series closed after EXQ-028 (2026-03-19). Governance cycle applied 7 decisions, V3-pending gate lifted.
• V3 active. ~96 experiments run (through EXQ-096a). SDs 004–010 implemented. SD-010 (harm stream separation) unblocked the prior FAIL cluster: EXQ-056c/058b/059c all PASS.
• SD-011 is the current bottleneck. Dual nociceptive streams (z_harm_s + z_harm_a) are required for the SD-003 counterfactual redesign. EXQ-093/094 confirmed that HarmBridge(z_world → z_harm) has bridge_r2=0 (architectural impossibility by SD-010 design). ~10 experiments are blocked pending SD-011. Design doc: sd_011_dual_nociceptive_streams.md.
• SD-012 registered. Homeostatic drive modulation for z_goal seeding — required for EXQ-085+ (wanting/liking experiments) and for any goal-directed behavior validation. Design doc: sd_012_homeostatic_drive.md.
• New claims registered (2026-03-24/25): INV-032–038 (approach/avoidance symmetry, epistemic self-monitoring, goal maintenance, state definition, stored/active distinction, EVR pattern). ARC-030–035. SD-011/012. MECH-112–134.
• Currently queued: EXQ-074b (MECH-112/117 wanting/liking, supersedes EXQ-074) and EXQ-076b (MECH-116/ARC-032 goal conditioning, supersedes EXQ-076).
• MECH-124 diagnostic: When reviewing EXQ-074b/076b results, check whether z_goal salience is competitive with harm salience — if not, this is a V4 early risk indicator (consolidation-mediated option-space contraction).
• 0 pending review as of 2026-03-25 (all experiments discussed and marked in review_tracker.json).
• Phase gate: SD-011 implementation → re-run blocked experiments → governance cycle.
• Step 3.1 (Substrate Debt Resolution) is the current active step; SD-008/009/010 are done; SD-011/012 remain.

---

Status Snapshot (2026-03-20) — archived

• V2 complete. All three hard-stop criteria triggered after EXQ-028. Governance cycle run 2026-03-19: 7 decisions applied, V3-pending gate lifted, ARC-024 and SD-010 registered.
• V3 active. 73 experiments run (through EXQ-059). Substrate SDs 004/005/006/007 implemented. EXQ-030b PASS validated V3-form SD-003 attribution pipeline (attribution_gap=0.035, world_forward_r2=0.947). Current focus: SD-010 implementation (harm stream separation) to unblock ~10 pending FAILs.
• Q-020 adjudicated: ARC-007 strict (2026-03-16). HippocampalModule generates value-flat proposals; terrain sensitivity is a consequence of navigating residue-shaped z_world, not a separate hippocampal value signal.
• 17 pending FAILs awaiting review (generated 2026-03-20). Root cause cluster: fused z_world containing harm signal → SD-010 substrate debt.
• REE_assembly is the canonical governance + specification repo. Current V3 roadmap is in §REE-v3 below; v2_v3_transition_roadmap.md is now historical.

---

Status Snapshot (2026-02-28) — archived

• REE_assembly is the canonical governance + specification repo.
• ree-v1-minimal has served as the qualification harness: 8 genuine experiments completed (EXQ-000 through EXQ-007), 4 PASS, 4 informative FAIL (substrate-limited).
• Substrate debt items SD-001, SD-002, SD-003 registered. V1 has reached saturation for the claims it was designed to test. Further V1 runs (EXQ-008/009) complete the current evidence cycle; extended-seed reruns (EXQ-010–013) are low-priority confidence accumulation.
• V2 cutover gates passed 2026-02-18 but cutover was deferred. Correct decision: V2 specification needs redesigning (Step 2.0, below) before implementation begins.
• JEPA integration guidance remains convergence-first: source-method details live in REE_convergence; REE_assembly keeps REE-first canonical contracts and adjudication outputs.

---

Roadmap Discipline

Each step must be completed in sequence. Before starting the next step:

1. Record what was learned (update GOVERNANCE_STATE.md, substrate debt register, any affected claims).
2. Update this roadmap to reflect that learning — revise subsequent steps if the evidence changes what they should be.
3. Make the update-roadmap action explicit in the exit criteria of every step.

The roadmap is not a fixed plan; it is a living document that is deliberately updated at each step boundary. Steps may be added, split, or reordered as understanding grows.

V3 Completion Gate For First Paper (2026-03-31 clarification)

This section records the current planning boundary for when ree-v3 should be considered “complete enough” for the first real paper.

Paper-1 target claim

REE should support a narrow, defensible claim:

• A waking, single-agent REE substrate can learn stable self/world attribution, harm avoidance, and genuine goal-directed behavior from architectural separation rather than a monolithic reward objective.

Must-pass gates

• Post-SD-011 SD-003 works in the current architecture. Counterfactual attribution must succeed on the sensory-discriminative harm stream (z_harm_s), not only on the older z_world formulation.
• Harm/attribution substrate is stable. SD-005, SD-010, SD-011, and ARC-033 must be reliable enough to function as substrate rather than as active rescue work.
• Goal-directed behavior is behaviorally real. SD-012 and MECH-112 must produce a genuine GOAL_PRESENT vs GOAL_ABSENT behavioral lift, not just z_goal seeding.
• Approach and avoidance compete in one selector. ARC-030 must be demonstrated as dual evaluation of the same candidate trajectories by harm and goal channels inside a shared commitment process.
• Results survive reruns. Matched-seed reruns and at least one task variant must confirm the core behavior is not a one-task artifact.
• Governance state is clean. Review tracking, claim status, and experiment indexing must remain aligned enough for the evidence story to be legible to an external reader.

Two-tier V3 completion (2026-04-02 clarification)

V3 completion has two levels with distinct gates:

V3 first-paper gate (sufficient for Paper 1 claim):

• Habit-system goal-directed behavior demonstrated: SD-012 activates approach drive; EXQ-182a oracle confirms the environment is near-optimal for habit-level policy; goal-lift experiment (EXQ-074e successor) shows GOAL_PRESENT > GOAL_ABSENT behavioral lift with ARC-030 harm/goal competition in one selector.

V3 full completion gate (required before V4 entry):

• HippocampalModule multi-step trajectory planning validated. This is a V4 prerequisite, not merely a V4 feature. V4’s social extension (“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. One-step greedy cannot reach this: it can approach its own resources but cannot plan paths that sustain another’s joy or reduce another’s sorrow over multi-step trajectories. The VTA/hippocampal system (MECH-163) must be validated in V3 to provide the planning substrate that V4 social cognition will depend on.
• All V3 first-paper gates passed.

See MECH-163 (dual goal-directed systems: habit vs hippocampally-planned).

Explicit non-blockers for V3 first-paper completion

• Sleep/offline consolidation mechanisms.
• Integrated self/other social modelling.
• Full language integration; simple future communication primitives are enough.
• nth-order multi-agent ethics tests.
• Full computational-psychiatry coverage.

Deferred to V4/V5 / later papers (requires V3 full completion gates first)

• Sleep-like consolidation as a load-bearing mechanism.
• Social coupling and other-modelling inside core substrate — specifically requires hippocampal multi-step planning (MECH-163 VTA/planned system) from V3 full gate.
• Structured communication between agents.
• Emergent ethical behavior in multi-agent settings.
• Stronger psychiatric modelling beyond single-agent perturbation analogs.
• INV-043 (caregiver requirement) — requires multi-agent substrate with modelled caregiving. V3 cannot test whether ethical capacity is motivationally activated (vs merely architecturally present). This is a first-class V4 research question.
• MECH-158 (love-exclusion failure mode) — requires developmental multi-agent substrate to test whether absence of love-experience collapses ethical motivation.
• MECH-159 (intergenerational moral progress) — requires multi-generation agent infrastructure. V5 scope or later.

The distinction between V3 and V4 is not only scope but epistemological level: V3 tests whether the machinery works. V4 tests whether the machinery develops correctly. The ARC-043 stack makes this precise: V3 exercises Layers 6-9; V4 must exercise Layers 2-5 dynamically, with caregiving, developmental phases, and social residue.

Deployment constraint: V3 may be used as a sandboxed scientific substrate, but serious capability scale-up or external-world connectivity is deferred until V4 social/developmental completion. Language alone is not treated as sufficient safety. See docs/governance/deployment_gating.md.

---

Phase Definitions

REE-v1 ✓ (completed purpose: qualification baseline)

Primary role: validate whether proposed mechanisms produce expected directional effects under controlled conditions.

Outcome: useful for signal discovery and contract hardening. Not sufficient as final architecture target due to stress-lane conflicts, limited environment breadth, and accumulated substrate debt (SD-001, SD-002, SD-003). Four genuine PASSes confirm core signal structure; four informative FAILs confirm substrate resolution limits, not architecture failures.

Post-V1 learning incorporated into roadmap:

• E2/hippocampus conflation (SD-001) prevents clean mechanistic isolation
• E1/E2 mutual constitution (SD-002) reframes timescale-separation interpretation
• E2 self-attribution substrate (SD-003) is an unmet V2 design requirement
• V1 stateless grid cannot surface persistent agent causal footprint
• MECH-057 (control completion) requires multi-step environment with genuine commitment pressure; needs richer substrate before further testing

---

REE-v2 ✓ (completed purpose: V2 qualification)

V2 series closed after EXQ-028 (2026-03-19). All three hard-stop criteria met. Governance cycle completed 2026-03-19.

Step 2.0 — V2 Redesign ✓

Primary role: Produce an updated V2 specification that incorporates V1 learning. The original V2 spec (representation-interface contract) remains in scope, but the design must now account for SD-001/002/003 and explicitly address what V2 must deliver to make future self-attribution and self-modelling experiments possible.

In-scope:

• Redesign V2 architecture to resolve SD-001: E2 separated as pure transition MLP, HippocampalModule created as distinct component of E3 complex
• Revise V2 entry criteria to include SD-003 requirements (see below)
• Produce first-pass V2 implementation spec: subsystem boundaries, required metrics, failure gates
• Define the persistent-causal-footprint environment requirement (SD-003)
• Capture mutual constitution (SD-002) in architecture documents

Out-of-scope:

• Implementation (code changes happen in Step 2.1 onward)
• New V1 experiments beyond EXQ-008/009

Exit criteria:

• Updated V2 implementation spec document exists with SD-001/002/003 addressed
• V2 entry criteria revised (see below)
• This roadmap updated with Step 2.1–2.5 refined based on redesign output
• GOVERNANCE_STATE.md SD-003 entry complete ✓ (done 2026-02-28)

Outcome: ✓ Complete. V2 spec produced (docs/architecture/ree_v2_spec.md). SD-001/002/003 addressed in design. Steps 2.1–2.5 scoped.

---

Step 2.1 — E2 Separation (SD-001 resolution) ✓

Primary role: Refactor E2 into a pure fast transition model. Create HippocampalModule as a distinct component of the E3 complex. Close SD-001.

In-scope:

• E2FastPredictor → pure forward predictor: forward(z, a) → z_next (cerebellum-like)
• HippocampalModule (new) → trajectory proposal by navigating affective terrain; not by running transition predictions
• Counterfactual E2 querying made architecturally possible: e2.forward(z, a_cf)
• SD-001 closed in GOVERNANCE_STATE.md

Out-of-scope:

• Full E3 complex redesign (Step 2.2+)
• Self-attribution experiments (Step 2.4)

Exit criteria:

• E2 callable independently with arbitrary action input
• HippocampalModule exists as separate class
• Existing EXQ PASS results replicated on refactored substrate (parity check)
• Roadmap updated with any discoveries

Outcome: ✓ Complete. E2 is a pure fast transition MLP on z_self. HippocampalModule created as distinct E3 component. SD-001 closed.

---

Step 2.2 — Representation Interface Contract ✓

Primary role: Lock stable representation-interface contract for sensing adapters and E1/E2 latent prediction. This is the original primary V2 role, now sequenced after E2 separation.

In-scope:

• Sensor adapters mapped to JEPA-like context/target latent interfaces
• E1/E2 representation-reference integration contract (IMPL-022)
• Stable output streams for latent prediction error and uncertainty
• Run-pack/adapter-signal compliance and calibration metrics

Out-of-scope:

• Full control-plane completion
• Hippocampal/E3 commitment architecture
• Full ethical arbitration dynamics

Exit criteria:

• Representation-interface contract stable across qualification and stress lanes
• Uncertainty/error streams calibrated and non-gamed across distribution shifts
• No unresolved adapter contract drift
• Roadmap updated with any discoveries

Outcome: ✓ Complete. MECH-059 PASS confirmed E1 precision and E3 confidence are structurally independent. Representation interface stable.

---

Step 2.3 — Persistent Causal Environment ✓

Primary role: Upgrade the environment substrate so that persistent agent causal footprint is present — actions at step N affect the landscape at step N+k in ways that require disambiguation from independent environment change. This is the prerequisite for SD-003 experiments.

In-scope:

• Environment design where agent-caused and environment-caused transitions are structurally distinct and must be separated for correct attribution
• Validation that E2 (now pure transition model from Step 2.1) can be queried counterfactually against this environment

Out-of-scope:

• Full self-attribution claim testing (Step 2.4)

Exit criteria:

• Environment exists with persistent agent causal footprint
• Baseline experiments confirm agent-caused / environment-caused distinguishability
• Roadmap updated with any discoveries

Outcome: ✓ Complete. CausalGridWorld implemented with persistent agent causal footprint. EXQ-018 PASS confirmed agent-caused/environment-caused distinguishability.

---

Step 2.4 — Self-Attribution Substrate ✓ (V2 partial; V3 form validated)

Primary role: Implement and test counterfactual E2 querying for self-modelling. First genuine experiments on self-attribution claims. Close SD-003.

In-scope:

• Counterfactual E2 queries integrated into agent decision loop
• First genuine experiments isolating agent-caused vs environment-caused harm
• Self-attribution claim coverage (claims to be identified during Step 2.0 redesign)
• SD-003 closed in GOVERNANCE_STATE.md

Exit criteria:

• At least one genuine PASS on a self-attribution claim
• E2 counterfactual querying demonstrated experimentally
• Roadmap updated with any discoveries

Outcome: ✓ Partial / hard stop. EXQ-027 FAIL triggered V2 hard stop: E2 cannot discriminate agent-caused harm in z_gamma; SD-005 substrate required. V2 self-attribution experiments concluded. V3-form SD-003 validated at EXQ-030b PASS (2026-03-18): attribution_gap=0.035, world_forward_r2=0.947. Full SD-003 achieved on V3 substrate with z_world separation.

---

Step 2.5 — V2 Qualification ✓

Primary role: Genuine experiment coverage across core V2 claims. Sufficient evidence to make V3 entry decision.

Exit criteria:

• Representation interface stable across qualification and stress lanes (from Step 2.2)
• Self-attribution substrate tested (from Step 2.4)
• Governance confidence above provisional thresholds for core V2 claims
• Roadmap updated with V1+V2 learnings before V3 begins
• V3 entry decision made explicitly

Outcome: ✓ Complete. V2 series closed after EXQ-028. 15 genuine V2 experiments (EXQ-014–028): 5 structural-separation PASSes, 9 FAILs (all substrate-limited by z_gamma conflation or SD-004 absence). Governance cycle 2026-03-19: 7 decisions applied. V3 entry formally made.

---

REE-v3 ← current phase (control completion + full attribution)

Primary role: Implement full attribution pipeline, control-plane heartbeat architecture, and E3 commitment/accountability on the z_self/z_world split substrate (SD-004/005).

Step 3.0 — V3 Substrate Implementation ✓

SD-004 (E2 action objects; HippocampalModule navigates action-object space O), SD-005 (z_gamma → z_self + z_world split), SD-006 (asynchronous multi-rate heartbeat, time-multiplexed phase 1), SD-007 (ReafferencePredictor for perspective-corrected z_world). Q-020 adjudicated: ARC-007 strict (2026-03-16). CausalGridWorld extended to V3 environment. EXQ-030b PASS: SD-003 attribution pipeline validated.

Step 3.1 — Substrate Debt Resolution ← current step

In-scope:

• SD-008: alpha_world ≥ 0.9 in LatentStackConfig ✓ (validated EXQ-040)
• SD-009: event-contrastive CE auxiliary loss for z_world encoder ✓ (EXQ-020 PASS)
• SD-010: harm stream separation ✓ (EXQ-056c/058b/059c PASS)
• SD-011: dual nociceptive streams (z_harm_s + z_harm_a) ← current focus
• SD-012: homeostatic drive modulation for z_goal seeding ← next

Exit criteria:

• SD-011 implemented and EXQ-093/094 successors run on dual-stream substrate
• SD-003 counterfactual redesigned for z_harm_s pipeline and validated
• SD-012 implemented; EXQ-085 successors (wanting/liking) runnable with functional z_goal
• Pending FAIL cluster (~10 experiments) reviewed after SD-011 implementation

Step 3.2 — V3 Claim Qualification

In-scope:

• ARC-016: E3-derived dynamic precision + end-to-end commitment→behavior behavioral distinction
• MECH-025: action-doing mode probe on V3 substrate
• MECH-057b: completion gate retest
• Q-007: valence-precision interaction
• SD-006: multi-rate loop validation at scale (ARC-023, MECH-089–093)
• MECH-090: beta-gated policy propagation
• ARC-024: harm attribution with SD-010 substrate
• Full V3-EXQ series (V3-EXQ-001 through V3-EXQ-010 as designed in transition roadmap)

Exit criteria:

• Governance confidence above provisional thresholds for core V3 control claims
• V3-pending claims adjudicated (ARC-007, ARC-016, ARC-018, MECH-025, MECH-033, Q-007)
• Roadmap updated with V1+V2+V3 learnings before V4 begins

Step 3.3 — V3 Governance Cycle and V4 Entry Decision

Exit criteria to start V4:

• Robust separation of exploratory simulation vs committed learning
• Stable behaviour under adversarial trajectory pressure
• Governance confidence above provisional thresholds for core control claims
• V4 entry decision made explicitly

---

REE-v4 (later: social and institutional complexity)

Primary role: Scale to richer multi-agent coupling, language-mediated coordination, and institutional constraints.

Exit criteria: To be defined during V3, informed by what V3 delivers.

---

V2 Entry Criteria (revised 2026-02-28)

Original criteria (all still required):

• Representation-interface contract stable across qualification and stress lanes
• Uncertainty/error streams calibrated and non-gamed across distribution shifts
• No unresolved adapter contract drift

Added from V1 learning:

• SD-001 resolved: E2 implemented as pure isolatable transition model; HippocampalModule exists as separate component of E3 complex
• SD-003 requirement: Counterfactual E2 querying architecturally possible (e2.forward(z, a_counterfactual) callable independently)
• Persistent causal environment: environment substrate provides persistent agent causal footprint, enabling agent-caused vs environment-caused transition disambiguation

---

Repository Roles

• REE_assembly: canonical claims, architecture docs, evidence matrix, governance outputs.
• ree-v3: primary qualification lane for V3 experiments and claim coverage. Default branch: main. Results go to REE_assembly/evidence/experiments/.
• ree-v2: transitional reference. V2 series complete (EXQ-014–028). No new experiments.
• ree-v1-minimal: legacy baseline/reference harness. No new mechanism development.
• ree-experiments-lab: ARCHIVED 2026-02-26. Synthetic scaffolding only; do not use.

---

Immediate Work Queue (This Cycle)

Current step: SD-003 Successor Validation + First-Paper Gate (as of 2026-04-18)

SD-004 through SD-023 all implemented. ARC-033, MECH-090 (bistable + Layer 1 trajectory stepping), MECH-091 Layer 2 urgency interrupt, MECH-120, MECH-203/204, MECH-205, MECH-216 implemented. SD-003 superseded 2026-04-18 by MECH-256 + SD-029 + MECH-257. Governance cycle 2026-04-18 applied 2 hold_pending_v3_substrate (SD-014, SD-023). 0 pending review. Second Hetzner worker (ree-cloud-2) onboarded.

1. V3-EXQ-433 (SD-029 event-conditioned single-pass comparator on z_harm_s — decisive test of the new self-attribution topology after SD-003 supersession; next-up priority=60).
2. V3-ONBOARD-smoke-ree-cloud-2 (second Hetzner worker calibration smoke).
3. V3-EXQ-321c (MECH-090 bistable vs legacy gate hold rate, spike-aligned E3-tick fix).
4. V3-EXQ-325b (SD-021 descending pain modulation retest, E2 world-forward training fix).
5. V3-EXQ-330a (SD-013 contrastive counterfactual retest).
6. V3-EXQ-418b (SD-016+SD-017 context-conditioned action: SHY fix + terrain_loss).
7. V3-EXQ-326 / V3-EXQ-326a (SD-015 wanting-gradient nav and MECH-229 behavioral dissociation fix).
8. V3-EXQ-434 / V3-EXQ-406b (INV-053 depression attractor replication; 5-seed LONG_HORIZON characterisation).
9. V3-EXQ-435 (INV-054 phase-transition recovery, sustained-crossing criterion, supersedes EXQ-278).
10. V3-EXQ-436 (SD-017 sleep phase ablation redesign with context-conditioned harm threshold, supersedes EXQ-242).
11. V3-EXQ-429b (INV-044 Bayesian prior-before-posterior; SWS-ordered vs REM-only).
12. V3-EXQ-407 (MECH-231 E2 short-horizon efference-copy discriminative pair).

---

ARC-057 Environment-Complexity Gate (2026-04-14)

Status: PARKED. ARC-057 (curiosity-approach emergence) cannot be faithfully tested in the current CausalGridWorld. The mechanism requires an environment where representational expansion at a location captures genuinely additional information – near-fractal complexity where zooming in reveals more structure. Grid cells are informationally flat: a cell is a cell. See docs/architecture/hippocampal_valence_asymmetry.md.

What this blocks: Faithful testing of approach-via-representational-expansion (MECH-232), the valence encoding asymmetry (MECH-233), and the curiosity-approach emergence (ARC-057).

What this does NOT block: The threat/avoidance side (harm residue field, BLA-pathway logic) remains fully testable. All existing V3 architecture work continues. The claims are registered and constrain future design even without experiments.

Proxy mechanism policy

Any experiment testing approach behavior in the grid world will necessarily use a proxy mechanism for the approach signal (e.g., explicit wanting gradient, DA-modulated place priority, or direct goal-location bias). This is acceptable for testing downstream architecture (commitment gating, trajectory selection, E3 evaluation) but carries a contamination risk:

Results obtained with a proxy approach mechanism may not generalize to an ARC-057-enabled agent. The proxy is a commanded gradient; ARC-057 is emergent from representational expansion + curiosity. The downstream architecture may develop implicit dependencies on proxy properties (e.g., gradient smoothness, signal strength, spatial extent) that the real mechanism would not provide.

Tagging requirements

1. Any experiment that tests approach behavior must declare in its docstring and manifest notes which proxy mechanism it uses for the approach signal.
2. Tag: experiments using a proxy approach mechanism should include approach_proxy in their tags. This enables future filtering when ARC-057 becomes testable.
3. Evidence interpretation: PASS results for claims that depend on approach behavior should carry an evidence_quality_note stating the proxy was used and results may not transfer to ARC-057-enabled substrate.
4. Re-validation queue: when a richer environment becomes available, all approach_proxy tagged experiments form the re-validation backlog.

What “richer environment” requires

• More sensory input channels (visual texture, object features, spatial micro-structure)
• Larger latent spaces to encode location-dependent detail
• Significantly more compute for training
• Location-dependent information density (some areas genuinely have more to discover)

This is a V5+ concern. V3 and V4 proceed with proxy mechanisms and honest tagging.

---

Open Questions

• SD-010 HarmEncoder architecture: should z_harm be a separate stream alongside z_world and z_self, or should it route through z_world after SD-010? The l_space.md architecture suggests a four-stream model (z_self, z_world, z_beta, z_harm).
• ARC-016 E3-derived precision: EXQ-038 FAIL — root cause (precision invariance) needs analysis before designing the next precision-regime experiment.
• SD-006 phase 2: time-multiplexed multi-rate is phase 1; true asynchronous execution (thread-based or event-loop) is still open. HTA (hierarchical temporal abstraction) is the recommended direction but not yet designed.
• ARC-057 environment substrate: what is the minimum environment complexity that would enable faithful testing of curiosity-approach emergence? Is a continuous 2D world with procedural texture sufficient, or does it require full 3D/visual input? See ARC-057 gate section above.

---

Related Claims

• IMPL-008, IMPL-020, IMPL-021, IMPL-022
• MECH-057, MECH-058, MECH-059, MECH-060, MECH-063
• ARC-024, MECH-069, MECH-070, MECH-089, MECH-090, MECH-100, MECH-101, MECH-102
• SD-010 (harm stream separation)

References

• docs/architecture/jepa_ree_hybrid_diagram_spec.md
• docs/architecture/jepa_e1e2_integration_contract.md
• evidence/GOVERNANCE_STATE.md (substrate debt register: SD-001, SD-002, SD-003)
• evidence/experiments/claim_evidence.v1.json
• evidence/experiments/promotion_demotion_recommendations.md
• evidence/planning/CUTOVER_REE_V2_READINESS.md