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 (2026-04-06)

  • 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.

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: 3.1 -> 3.2 transition — First-Paper Gate Experiments

SD-008/009/010/011/012/014 all implemented. ARC-028+MECH-105 hippocampal-BetaGate coupling implemented. ARC-033 ResidualHarmForward in ree_core. 0 pending review (clean). EXQ-125 currently running (ARC-029 committed mode pair, DLAPTOP-4.local).

  1. EXQ-125 (ARC-029 committed vs ablated mode harm pair; running): validate committed vs uncommitted mode harm outcomes in stable and volatile environments.
  2. EXQ-074e (MECH-112/117 wanting/liking dissociation; supersedes EXQ-074d): tests z_goal_latent vs benefit_eval_head. MECH-124 diagnostic: z_goal salience vs harm salience as V4 risk indicator (consolidation-mediated option-space contraction).
  3. EXQ-076e (MECH-116/ARC-032 E1 LSTM goal conditioning; supersedes EXQ-076c): frontal working memory for goal representation over multi-step episodes.
  4. EXQ-195 (SD-003 z_harm_s counterfactual, post-SD-011): critical re-validation of SD-003 using ResidualHarmForward (ARC-033). Pipeline: z_harm_s_cf = E2_harm_s(z_harm_s, a_cf). Do NOT use HarmBridge (bridge_r2=0).
  5. EXQ-223 (minimal vertebrate ablation, queued priority 5): PASS already confirmed 2026-04-03; queued for a fresh run to confirm repeatability.
  6. EXQ-184/196 (MECH-033 kernel chain, ARC-018 rollout viability): queued priority 4.

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.
  • 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

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.