Glossary

Claim Type: implementation_note
Scope: Repository terminology and canonical term definitions
Depends On: None
Status: stable
Claim ID: IMPL-001

  • REE: Reflective‑Ethical Engine.
  • E1: Deep predictor (long-horizon context model).
  • E2: Fast predictor (short-horizon reflex model).
  • E3: Trajectory selector (planning and commitment module).
  • L-space: Fused latent manifold stratified by prediction depth.
  • Rollout: Explicit multi-step trajectory sequence generated by hippocampal systems; used for candidate futures.
  • Forward prediction: Short-horizon conditional transition kernel (typically 1–2 steps) produced by E2; used to seed hippocampal rollouts.
  • Commit boundary: Authority transition where irreversible action dispatch or privilege-bearing durable-write eligibility begins for a specific commit_id; not a requirement to pause stream processing.
  • Pre-commit channel: Continuous rehearsal/simulation error stream used for search, gating, and arbitration; cannot directly perform durable responsibility-bearing writes.
  • Post-commit channel: Continuous realized-outcome/reafference stream joined to commit lineage and allowed to drive durable attribution/policy/residue updates.
  • Superseding commit: A new commit event that interrupts/re-routes a previously dispatched commit; extends lineage history without erasing accountability of the earlier commit.
  • Hazard class: Typed safety category attached to superseding commits, indicating why fast interruption authority was invoked.
  • Override scope: Bounded mitigation envelope for a superseding commit (for example halt, brake, evade, contain, safe-mode transition).
  • (\zeta): Selected trajectory (latent + hippocampal action rollout).
  • (\mathcal{F}): Reality constraint (computable proxy for VFE (Variational Free Energy)).
  • (M): Legacy ethical cost proxy (predicted degradation of self/other variables). Retained for traceability and evaluation; current canonical framing does not require an explicit ethical cost term (see docs/architecture/e3.md).
  • Residue (R): Persistent curvature term produced by ethical consequence history (legacy formulations tie this to (M)).
  • (\phi(z)): Residue dent field over latent space.
  • Precision (\alpha_k): Depth-indexed gain controlling error weighting / entropy.
  • Conceptual stack (ARC-043): The Layer 0-9 ordering of REE’s full architecture from epistemic ground through the learning loop. Layers 0-4 are the five axioms (cannot be updated by experience); Layer 5 is derived ethics; Layer 6 is REE itself; Layers 7-9 are actions, consequences, and learning. See docs/architecture/five_axioms_foundations.md.
  • Derived ethical objectives (INV-042): The nine objectives that follow necessarily from the five axioms: preserve minds; preserve future options; reduce unnecessary suffering; increase shared joy; maintain corrigibility; maintain truth-seeking; maintain the ability to love and be loved; maintain the shared world; maintain the possibility of future minds and future love. These are not design choices — they are what the axioms jointly require.
  • Associative manifold (MECH-154): E1 conceived as an indexed substrate supporting retrieval, traversal, and sequential planning — not merely a representation space. The manifold is addressable: items are indexed so that the system can navigate to, traverse across, and compose sequences over stored content. Implemented in cerebral cortex including parietal associative geometry.
  • External/internal coupling modes (MECH-157): Precision-routing modes that govern the balance between sensory-coupled perception and internally generative simulation. External mode: high sensory gain, low hippocampal drive; Internal mode: low sensory gain, high hippocampal drive. Mode switching is a control-plane operation. See docs/architecture/modes_of_cognition.md.
  • Caregiver requirement (INV-043): The architectural constraint that REE’s ethical capacity — though structurally present from initialisation — requires caregiving experience (being loved and modelled as loveable) to become motivationally active. Architecture is necessary but not sufficient for ethical behaviour. Testing this constraint requires multi-agent substrate (V4+). See docs/architecture/developmental_curriculum.md.
  • Love-exclusion failure mode (MECH-158): A developmental failure mode in which an agent acquires the concept that love exists but infers it does not apply to itself. The result is ethical motivation collapse: self-preservation and goal-pursuit without moral constraint. The agent may retain the cognitive structure of ethics while lacking the motivational activation that makes it action-guiding.
  • Intergenerational moral progress (MECH-159): The hypothesis that moral progress across agent generations requires active transmission — not mere environmental exposure. Each generation must receive ethical orientation through caregiving, or moral capacity re-emerges from scratch. Implies that multi-agent REE systems require designed caregiving continuity, not just architectural replication. See docs/architecture/social.md.

External Vocabulary Alignment (JEPA to REE)

Formal Alignment Glossary (IMPL-020)

Purpose: align terminology between latent predictive world-model research (JEPA framing) and REE so claims are structurally interoperable without forcing conceptual merger.

Canonical scope:

  • maintain only adjudicated JEPA-to-REE term translations needed by canonical REE docs and contracts,
  • keep source-specific and exploratory translation drafts in REE_convergence,
  • promote external translation changes back through convergence packets before canonical adoption.

Canonical term translation set

JEPA term REE term Canonical interpretation
Embedding / latent representation L-space Shared latent manifold where prediction and rollout preparation operate.
Encoder / context-target encoder E1 deep integrative predictor Slow cross-modal and cross-temporal integration path.
Latent predictor E2 fast predictive generator Short-horizon latent transition generator.
Target embedding Projected future L-state Latent future target used for predictive alignment.
Joint embedding alignment loss / latent prediction error L-state prediction error signal (PE routing input) Deviation signal for adaptation without requiring one global scalar objective.
Temporal/video JEPA context Multi-timescale predictive integration (E1+E2 coupling) Coupled short- and long-horizon trajectory modeling.
Action-conditioned prediction Efference/reafference loop Embodied prediction stream conditioned on intervention.
Self-model in latent manifold Self L-state modeling Viability-critical self representation.
Control mechanism (implicit gap in JEPA) E3 trajectory selector + control plane Explicit trajectory arbitration and gating under uncertainty.
Representation collapse avoidance / predictable invariants Viability constraint + shared modelability Stability requirements spanning single- and multi-agent conditions.
Multi-agent stability requirement Ethical stability constraint Preservation of shared predictability as a viability condition.

Divergence points

  • JEPA is typically objective-centric; REE is constraint-centric.
  • JEPA does not require ethical modeling; REE treats ethical stability as structural viability.
  • JEPA does not define control-plane arbitration; REE treats control arbitration as irreducible.

Minimal interoperability principle

A JEPA-like world-model design can serve as a REE E1/E2 representational reference architecture. JEPA itself remains an external project and is not a component inside REE. REE contributes substrate ownership, control-plane architecture, E3 commitment, self-impact attribution, trajectory constraints, and multi-agent stability modeling.

JEPA scope boundary (Consolidation Selectivity, 2026-02-26)

Empirical compact modelling of macaque V4 (Cowley et al. 2023; MECH-068) clarifies the JEPA scope boundary within REE:

  • JEPA covers stage 1: the shared feature basis / early filter bank — E1 representation.
  • JEPA does not cover stage 2: the consolidation/gating step that determines behavioural selectivity — E3 trajectory gating and the control plane.

JEPA should be treated as a reference architecture for E1 representation only. The consolidation step — where trajectory preference, value weighting, and commitment identity emerge — is outside JEPA’s design scope. This is not a limitation of JEPA; it is simply a different architectural concern.

Consequence: integration work that attempts to use JEPA as the complete REE substrate (rather than as an E1 representation reference) is out of scope and should not proceed. JEPA-related convergence packets should be scoped to E1 representation only.

See docs/architecture/compact_consolidation_principle.md (MECH-068).

Wording and boundary policy

For canonical docs in REE_assembly:

  • use REE-first wording with optional JEPA translation when needed:
    • REE term (JEPA term)
  • keep REE schema keys stable in machine contracts.
  • keep extended source-level term tables in REE_convergence:
    • sources/jepa/glossary.md
    • translations/glossary_master.md
  • require convergence packet references for canonical JEPA terminology updates.

For source-intake and integration-playbook docs in REE_convergence:

  • JEPA-first wording is acceptable and often preferred for source fidelity.

Reference:

  • docs/notes/jepa_language_policy.md#impl-024
  • evidence/planning/CONVERGENCE_EXTRACTION_MAP.md
  • evidence/planning/REE_CONVERGENCE_INTERFACE.md

Open Questions

None noted in preserved sources.

  • IMPL-001
  • IMPL-020
  • IMPL-024
  • ARC-043 (conceptual stack)
  • INV-042 (derived ethical objectives)
  • MECH-154 (E1 as associative manifold)
  • MECH-157 (external/internal coupling modes)
  • INV-043 (caregiver requirement)
  • MECH-158 (love-exclusion failure mode)
  • MECH-159 (intergenerational moral progress)

References / Source Fragments

  • docs/processed/legacy_tree/docs/glossary.md
  • docs/thoughts/2026-02-13_jepa_ree_formal_alignment_glossary.md

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