JEPA to REE Hybrid Diagram Spec

Claim Type: implementation_note
Scope: Canonical diagram specification for JEPA↔REE architectural interoperability
Depends On: IMPL-020 (JEPA↔REE glossary), MECH-057 (control completion), ARC-001 (E1), ARC-002 (E2), ARC-003 (E3), ARC-005 (control plane), ARC-015 (responsibility flow)
Status: stable
Claim ID: IMPL-021

Purpose

Define a single, versioned diagram contract that shows how JEPA-style latent predictive modeling aligns with REE control-complete agency.

This spec is for:

design reviews,
evidence annotation,
and experiment-to-architecture traceability.

It is not a proof of equivalence between JEPA and REE.

Scope and Non-Goals

In scope:

representation alignment (embedding ↔ L-space),
predictive alignment (latent predictor ↔ E2),
control completion (E3 + control plane + responsibility loops),
evidence hooks for claim updates.

Out of scope:

implementation-specific tensor shapes,
optimizer details,
and claim promotion decisions.

Diagram Layers

The diagram MUST include five lanes (left-to-right flow, top-to-bottom control overlays):

Environment lane
- ENV.observation_stream
- ENV.action_consequences
- ENV.multi_agent_context
JEPA representational lane
- JEPA.context_encoder
- JEPA.target_encoder
- JEPA.latent_predictor
- JEPA.embedding_alignment_objective
REE predictive substrate lane
- REE.L_space
- REE.E1_deep_integrator
- REE.E2_fast_predictor
- REE.hippocampal_rollout_interface
REE control-completion lane
- REE.reafference_comparator
- REE.E3_trajectory_selector
- REE.control_plane_router
- REE.residue_viability_constraints
Governance/evidence lane
- EVID.claim_links
- EVID.literature_entries
- EVID.experiment_packs
- EVID.promotion_demotion_queue

Required Edges

The diagram MUST render at least these edge classes:

EDGE-001 observation -> JEPA.context_encoder
EDGE-002 context/target embeddings -> JEPA.latent_predictor
EDGE-003 JEPA predicted embedding -> REE.L_space
EDGE-004 REE.L_space -> REE.E1_deep_integrator
EDGE-005 REE.L_space -> REE.E2_fast_predictor
EDGE-006 REE.E2_fast_predictor -> REE.reafference_comparator
EDGE-007 action trace/efference copy -> REE.reafference_comparator
EDGE-008 REE.reafference_comparator -> REE.control_plane_router
EDGE-009 REE.control_plane_router -> REE.E3_trajectory_selector
EDGE-010 REE.residue_viability_constraints -> REE.E3_trajectory_selector
EDGE-011 REE.E3_trajectory_selector -> environment actions
EDGE-012 evidence outputs -> governance queue (EVID.*)

Overlay Requirements

Every rendered hybrid diagram SHOULD include three overlays:

timescale_overlay: marks fast (E2) vs slow (E1/residue) channels.
ownership_overlay: marks where self-attribution enters (reafference_comparator).
constraint_overlay: marks where trajectories are gated (control plane + E3 + residue).

Evidence Hooks

Each diagram node/edge should map to claim IDs and evidence paths.

Minimum required claim hooks:

MECH-057 (agentic control completion requirement)
Q-012 (falsifiability: stability without REE-like control)
ARC-015 (responsibility flow)
ARC-005 (control plane)

Minimum required evidence pointers:

evidence/literature/targeted_review_mech_057/**
evidence/experiments/claim_evidence.v1.json
evidence/experiments/conflicts.md
evidence/experiments/promotion_demotion_recommendations.md

Mermaid Contract (v1 Skeleton)

flowchart LR
  subgraph ENV[Environment]
    obs["ENV.observation_stream"]
    actc["ENV.action_consequences"]
    mac["ENV.multi_agent_context"]
  end

  subgraph JEPA[JEPA Representational Lane]
    jctx["JEPA.context_encoder"]
    jtgt["JEPA.target_encoder"]
    jpred["JEPA.latent_predictor"]
    jloss["JEPA.embedding_alignment_objective"]
  end

  subgraph REEPRED[REE Predictive Substrate]
    ls["REE.L_space"]
    e1["REE.E1_deep_integrator"]
    e2["REE.E2_fast_predictor"]
    hpc["REE.hippocampal_rollout_interface"]
  end

  subgraph REECTRL[REE Control Completion]
    rea["REE.reafference_comparator"]
    cp["REE.control_plane_router"]
    e3["REE.E3_trajectory_selector"]
    rv["REE.residue_viability_constraints"]
  end

  subgraph GOV[Evidence and Governance]
    cl["EVID.claim_links"]
    lit["EVID.literature_entries"]
    exp["EVID.experiment_packs"]
    pd["EVID.promotion_demotion_queue"]
  end

  obs --> jctx
  jctx --> jpred
  jtgt --> jpred
  jpred --> jloss
  jpred --> ls
  ls --> e1
  ls --> e2
  e2 --> rea
  rea --> cp
  cp --> e3
  rv --> e3
  e3 --> actc
  mac --> ls
  e1 --> hpc
  e2 --> hpc
  lit --> cl
  exp --> cl
  cl --> pd

Review Checklist

Before using this diagram in governance discussion:

all required lanes exist,
required edges are present and directionally consistent,
claim IDs are attached to control edges,
at least one literature and one experimental evidence link are attached,
unresolved divergences are called out explicitly (not hidden in labels).

IMPL-021
IMPL-020
MECH-057
Q-012
ARC-005
ARC-015