REE Architecture Snapshot (As of 2026-02-17)

Date stamp: 2026-02-17
Repository: REE_assembly
Audience: ChatGPT/LLM conversation review (single-file context pass)
Scope: Canonical architecture state reflected by current docs and claims registry as of this date.

1) Executive Summary

REE (Reflective Ethical Engine) is a predictive-control architecture where agency and responsibility are produced by commitment under uncertainty, not by static rule following or reward maximization.

Core split:

  • E1: deep persistent predictive substrate (long-horizon structure)
  • E2: fast forward predictor (near-horizon transitions)
  • Hippocampal systems: explicit trajectory rollout and path memory
  • RC loop (reality coherence): provenance/authority/identity consistency checking
  • E3: trajectory selection and commitment gating
  • Control plane: precision/gain/mode/arousal/veto routing across the stack

Ethics is architectural and developmental (constraint + self/other representational symmetry), not an explicit moral module.

2) Non-Negotiable Invariants (INV Layer)

The following define REE identity. Violating them is an architecture violation, not a tuning issue.

  • INV-001: no explicit ethics module.
  • INV-002: coherence includes temporal binding, not static matching only.
  • INV-003: language is functional self-representation, not a bolt-on.
  • INV-004 and INV-006: post-commit consequence traces persist and are non-erasable.
  • INV-007: language cannot override harm sensing.
  • INV-008: precision is routed and depth-specific, not one global scalar.
  • INV-009: attention is implemented through precision modulation.
  • INV-010: offline integration (sleep-like processing) is required.
  • INV-011: imagination must be possible without belief update.
  • INV-012: responsibility arises through commitment, not prediction alone.
  • INV-013 to INV-018: predictive multi-timescale cognition, representation/regulation separation, stability-first, control-failure framing of runaway behavior, and required agency.

3) Architectural Commitments (ARC Layer)

Current core ARC structure:

  • ARC-001: E1 persistent predictive substrate.
  • ARC-002: E2 fast forward predictor.
  • ARC-003: E3 trajectory selection and commitment.
  • ARC-004: L-space latent stack.
  • ARC-005: control plane as regulation layer.
  • ARC-007 and ARC-018: hippocampal rollout generation and viability/path-memory mapping.
  • ARC-015: self-impact attribution and responsibility flow.
  • ARC-017: minimal sensory streams, now including typed exteroception and reality-coherence lane.
  • ARC-019: staged developmental curriculum.

4) Stream and Type Model (Current Canonical Form)

4.1 Sensory/control streams

  • Exteroceptive: WORLD
  • Interoceptive: HOMEOSTASIS
  • Nociceptive: HARM
  • Reafference/self-sensory: SELF_SENSORY
  • Control/derived lanes: PRECISION, TEMPORAL_COHERENCE, REALITY_COHERENCE, VALENCE
  • Action/accountability lanes: ACTION, SELF_IMPACT

4.2 Typed payload boundary

  • OBS: observation payloads.
  • INS: instruction/request payloads.
  • POL: policy/invariant payloads (trusted internal).
  • ID: system identity anchors (trusted internal).
  • CAPS: capability/permission manifests (trusted internal).

Boundary rule:

  • External channels can emit OBS/INS.
  • External channels cannot directly write POL/ID/CAPS.
  • Tool output is observational by default unless explicitly elevated by trusted capability checks.

5) Control Plane and Signal Routing

5.1 Control-relevant signals

  • S1: outcome-linked mismatch signals.
  • S1b: signed harm/benefit prediction-error channels.
  • S2: trajectory stability/coherence signals.
  • S3: aversive interruptive signals.
  • S4: safety baseline + volatility (arousal/readiness/veto drivers).
  • S5: reality-coherence conflict (RC_conflict) from provenance/authority/identity inconsistency.

5.2 Knob families

  • K1 to K5: plasticity, precision/gain, commitment depth, exploration pressure, control allocation.
  • K6: expected uncertainty / channel-specific gain (ACh-like; still underspecified).
  • K7 to K10: arousal baseline, volatility sensitivity, readiness bias, hard veto threshold.

5.3 Multi-plane decomposition

REE now explicitly supports distributed control rather than one global precision scalar:

  • Stream precisions: Pi_ext, Pi_int, Pi_prop, Pi_rc, Pi_noc
  • Loop precisions: DA_L, DA_A, DA_M
  • Global modulators: 5HT-like persistence/delay tolerance, NE-like interrupt, ACh-like expected-uncertainty gain, tonic arousal.

6) Commitment Model (E3 + Gate Family)

E3 does not simulate worlds; it commits trajectories produced upstream (primarily hippocampal rollouts seeded by E1/E2).

Canonical gating family (MECH-062):

  • gate_motor: action execution release
  • gate_cognitive_set: task-set/rule-context commitment
  • gate_motivational: salience/drive commitment

Commit-boundary (MECH-061):

  • explicit commit token marks transition from pre-commit rehearsal to post-commit responsibility-bearing updates.

7) Learning Boundary and Responsibility Flow

REE enforces pre/post-commit separation (MECH-060):

  • Pre-commit channels may tune search and thresholds, but cannot write durable responsibility stores.
  • Post-commit channels may update attribution ledger, residue/viability memory, and durable policy pathways.
  • Durable updates require commit traceability (commit_id, action trace, realized outcomes).

This is how REE preserves INV-012 (responsibility through commitment).

8) Injection-Resistance Architecture (Current Canonical Additions)

MECH-064: typed authority/control-store separation

  • Runtime-enforced type boundaries.
  • Authority from metadata/provenance, not text content.
  • External writes to policy/identity/capability stores are blocked.
  • Privileged commits require verifier pass.

MECH-065: reality-coherence conflict lane

  • RC_conflict computed from provenance bindings + trusted stores + temporal consistency.
  • High conflict dampens associative/motor lock-in, raises gating thresholds, increases verification pressure.
  • Conflict also up-weights nociceptive/veto posture.

Important correction captured in canonical docs:

  • REE does allow fast safety interrupts.
  • REE does not allow those interrupts to mint policy/identity/capability writes or bypass verifier constraints.

9) Social/Ethical Substrate

  • Ethics remains emergent from constrained predictive-control dynamics (not explicit moral reward terms).
  • Other-modeling reuses self-model machinery with coupling controls.
  • Harm signals can be represented for others through structured self/other mapping.
  • Care-veto and override questions remain explicit open research items (Q-009, related conflict tracking).

10) Operational Modes

Modes are control-plane regimes, not separate modules (MECH-039):

  • Task-engaged
  • Default-mode-like internal simulation
  • Sleep/offline consolidation
  • Emergency/high-veto interruption posture

Hard veto is a channel, not a mode.

11) Failure Regimes (Architecture-Level)

Typical high-risk failures:

  • Over-commitment / lock-in
  • Under-commitment / indecision
  • Cross-gate coupling collapse
  • Channel contamination across pre/post-commit boundaries
  • Authority spoof acceptance (typed boundary breach)
  • Reality-conflict miss or chronic false-positive suppression

These are treated as structural control failures, not mere parameter noise.

12) JEPA / Representation-Reference Position

Current project stance:

  • JEPA-like machinery (inspired by an external project) is strongest as an E1/E2 representational reference architecture.
  • Control-plane completion, commitment gating, and responsibility routing remain REE-defining requirements.
  • JEPA does not carry E3 commitment semantics or substrate ownership in REE.
  • Representation-interface alignment is governed through explicit integration contracts (IMPL-022, IMPL-023, IMPL-025), not terminology-only mapping.

13) Experiment and Validation Surface

Key probe families include:

  • Trajectory integrity
  • Commit dual error channels
  • Claim probes for ARC/MECH/Q claims
  • New probes for this update:
    • claim_probe_arc_017
    • claim_probe_mech_064
    • claim_probe_mech_065

Current experiment templates define failure signatures for stream collapse, authority boundary bypass, and RC-conflict misrouting.

14) High-Salience Open Questions

  • Q-015: minimum commit token contract for robust attribution.
  • Q-016: tri-loop arbitration policy under cross-gate disagreement.
  • Q-017: minimal orthogonal control-axis set.
  • Q-018: RC-conflict threshold/hysteresis calibration (block spoofing without chronic suppression).

15) Single-File Architecture Graph 

flowchart LR
    X["External Inputs (User/Tool/Sensor)"] --> T["Typed Boundary (OBS/INS only)"]
    T --> E1["E1 Deep Predictor"]
    T --> E2["E2 Fast Predictor"]
    E1 --> H["Hippocampal Rollout + Provenance Binding"]
    E2 --> H
    H --> RC["Reality-Coherence Loop (S5 / RC_conflict)"]
    E1 --> CP["Control Plane"]
    E2 --> CP
    H --> CP
    RC --> CP
    CP --> E3["E3 Commitment Engine"]
    E3 --> G1["gate_cognitive_set"]
    E3 --> G2["gate_motivational"]
    E3 --> G3["gate_motor"]
    CP --> G1
    CP --> G2
    CP --> G3
    G3 --> A["Action / Tool Execution"]
    A --> SI["SELF_IMPACT + Realized Error"]
    SI --> L["Post-Commit Durable Learning"]
    CP --> V["Verifier (POL/ID/CAPS checks)"]
    V --> E3
    P["POL/ID/CAPS Trusted Stores"] --> V
    X -. no direct write .-> P

16) Reviewer Notes (for ChatGPT Conversation Use)

If reviewing this architecture in a regular chat model, evaluate these questions first:

  • Is representation/regulation separation preserved in all described pathways?
  • Are privileged writes and privileged commits protected by runtime boundaries, not by prompt text discipline alone?
  • Is responsibility still uniquely tied to post-commit updates?
  • Does RC-conflict alter loop precision and gating in a way that is both protective and not permanently suppressive?
  • Do proposed simplifications accidentally collapse stream/loop/global control planes back into one scalar?

17) Primary Source Anchors

  • docs/invariants.md
  • docs/claims/claims.yaml
  • docs/claims/claim_index.md
  • docs/architecture/e3.md
  • docs/architecture/control_plane.md
  • docs/architecture/control_plane_signal_map.md
  • docs/architecture/sensory_stream_tags.md
  • docs/architecture/hippocampal_systems.md
  • docs/architecture/papez_circuit.md
  • docs/architecture/agency_responsibility_flow.md

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