JEPA-Like E1/E2 Representation Reference Contract

Claim Type: implementation_note
Scope: Interface contract for using JEPA-like world-model patterns as E1/E2 representational reference architecture (JEPA project remains external)
Depends On: IMPL-020, IMPL-021, ARC-001, ARC-002, ARC-005, ARC-015, MECH-057
Status: stable
Claim ID: IMPL-022

Purpose

Specify how JEPA-like patterns can be integrated into REE as representational references for:

  • E1 deep predictive representation path (slow, coherence-preserving latent structure),
  • E2 fast predictive representation path (short-horizon transition prediction).

This contract isolates the representation boundary so control-completion behavior remains explicit REE scope.

Decoupling declaration:

  • JEPA is an external project and not a component of REE.
  • REE may use JEPA-like representational patterns as reference points for E1/E2 interfaces.
  • JEPA is not treated as REE substrate ownership.
  • JEPA cannot carry E3 commitment, responsibility attribution, or control-plane authority semantics.

Canonical boundary note:

  • this document defines the REE-side integration contract only,
  • source-specific JEPA integration playbooks, exploratory mappings, and probe planning live in REE_convergence (sources/jepa/*) and enter here only via convergence packets.

Functional mapping note:

  • E1 and E2 should be treated as functional decomposition over JEPA-like reference streams, not necessarily separate physical modules.
  • Fast online predictor adaptation can realize E2 behavior.
  • Slow target/anchor dynamics and continuity constraints can realize E1 behavior.
  • The required separations are update-rate asymmetry, typed routing, and write-locus permissions, not total latent isolation.

Integration Boundary

Reference-profile side provides:

  • latent state encoding,
  • latent future prediction,
  • optional action-conditioned latent dynamics.

REE side retains ownership of:

  • commitment gating (E3),
  • control-plane arbitration and precision routing,
  • responsibility attribution/efference-reafference learning,
  • residue and viability constraints.

Required Interface Surfaces

Inputs into JEPA-like reference adapter

  • obs_t: current observation slice.
  • ctx_{t-k:t}: context window (optional memory-conditioned input).
  • a_t (optional): candidate action token/vector for action-conditioned predictions.
  • mode_tags (optional): REE regime hints (deliberative, reactive, reflective) for evaluation-only stratification.

Outputs from JEPA-like reference adapter

  • z_t: latent representation for current context.
  • z_hat_{t+1:t+h}: predicted latent futures for horizon h.
  • pe_latent: latent prediction deviation statistics (L-state deviation).
  • uncertainty_latent (optional): calibrated uncertainty/dispersion estimate over latent predictions.

JEPA-inspired signal map (adapter contract)

The adapter must map native JEPA training/runtime streams onto stable REE-facing keys:

REE contract key JEPA-native stream Extraction rule Status
z_t encoder/target-encoder latent tokens export post-normalization latent for the active context window direct
z_hat_{t+1:t+h} predictor output tokens export predictor outputs aligned to masked target blocks and rollout step index direct
pe_latent.mean latent prediction loss compute mean latent distance from z_hat to target latent (smooth_l1/equivalent) direct
pe_latent.p95 latent prediction residuals compute p95 over per-token latent residuals before global reduction derived
pe_latent.by_mask masked block residuals emit per-target-mask residual mean keyed by mask id derived
uncertainty_latent.dispersion multi-target/multi-step latent spread estimate dispersion across equivalent target predictions (same context, multiple masks/seeds) derived
uncertainty_latent.calibration_error predicted uncertainty vs realized residual expected calibration error between uncertainty estimate and realized pe_latent bins derived
trace.context_mask_ids mask sampler outputs emit exact context/target mask ids used for each prediction direct
trace.action_token (if action-conditioned) action-conditioned predictor input emit action token/vector hash + schema id for each rollout step direct

Notes:

  • pe_latent is already explicit in canonical JEPA implementations.
  • uncertainty_latent is partially implicit in canonical JEPA and must be adapter-derived unless the producer exposes a dedicated uncertainty head.
  • precision_latent is not guaranteed as an explicit calibrated JEPA output channel. The adapter may export internal precision proxies (for example, dispersion, ensemble disagreement, attention entropy, or action sensitivity) with provenance metadata; REE control-plane logic must compose the final confidence channel (uncertainty-derived precision).
  • Machine-readable adapter declaration for experiment packs: evidence/experiments/schemas/v1/jepa_adapter_signals.v1.json.

JEPA↔REE control signal mapping matrix (execution contract)

This table defines ownership and gating for control-plane-relevant channels when a JEPA-like reference profile backs E1/E2.

REE control function REE-facing channel JEPA source class JEPA extraction examples Final transform owner Required checks Canonical failure signatures
Representation mismatch feed (MECH-058) pe_latent explicit/derived smooth_l1 residual, p95, by_mask REE consumes; adapter exports only latent_residual_coverage_rate >= 0.95 contract:jepa_residual_stream_missing, mech058:timescale_separation_collapse
Confidence-channel synthesis (MECH-059) confidence_channel derived proxy bank uncertainty_latent.dispersion, ensemble disagreement, attention entropy, rollout inconsistency, action sensitivity REE control plane only calibration, separability, and completeness checks mech059:calibration_slope_break, mech059:uncertainty_metric_gaming_detected, mech059:abstention_reliability_collapse
Pre-commit planning signal (MECH-060) pre_commit_error hybrid pe_latent + hypothetical action-conditioned rollout deltas REE control plane only pre_commit_error_signal_to_noise and leakage checks mech060:precommit_channel_contamination
Post-commit attribution signal (MECH-060) post_commit_error hybrid realized residuals + reafference delta + action trace REE attribution path post_commit_error_attribution_gain and attribution reliability checks mech060:postcommit_channel_contamination, mech060:attribution_reliability_break
Trajectory-first control support (MECH-056) trajectory_candidate_risk derived rollout consistency + commitment reversal diagnostics REE commitment layer trajectory-first checks under ablation mech056:trajectory_order_violation
Channel-isolation guard (MECH-060) cross_channel_leakage_rate REE diagnostic control-plane leakage estimator across pre/post channels REE control plane only cross_channel_leakage_rate <= threshold mech060:cross_channel_leakage_spike

Proxy bank declaration (adapter to control handoff)

When a JEPA-inspired control-routing profile is enabled, the adapter should export proxy_bank declarations for each active internal proxy used by REE confidence/channel logic. Each proxy entry should include:

  • proxy_id: stable key (uncertainty_dispersion, ensemble_disagreement, attention_entropy, rollout_inconsistency, action_sensitivity, other)
  • source_stream: source family (uncertainty_dispersion/ensemble_disagreement/attention_entropy/rollout_inconsistency/action_sensitivity/other)
  • extraction_method: short deterministic description of how the proxy is computed
  • normalization: scaling/bounding transform
  • window: temporal/statistical window used
  • calibration_target: what the proxy is calibrated against (latent_residual, commitment_reversal, attribution_gain, other)
  • provenance: adapter-internal source pointer (module name, tensor path, or equivalent)

Contract invariants

  • Output latents must remain numerically stable under repeated rollout calls.
  • Prediction deviation keys must be fixed-name numeric fields across runs.
  • JEPA-like adapter outputs must not directly commit actions; they are advisory inputs to REE control.
  • JEPA-like adapters may emit proxy signals, but must not emit final control-plane decisions or commitment actions.

Mapping Rules

  1. z_t feeds REE L-space ingest path.
  2. Multi-step z_hat feeds E2 short-horizon prediction interfaces and hippocampal rollout seeding.
  3. pe_latent feeds REE prediction-error routing (precision/eligibility inputs), not direct policy rewrite.
  4. Any action-conditioned JEPA-like output is treated as hypothetical until E3 commitment.
  5. Internal JEPA-like proxies are valid control inputs only after explicit REE-side calibration and attribution-safe routing.
  6. Shared latent producers are acceptable if E1/E2 contractual roles remain distinguishable by update-rate and routing policy.

Signed PE and precision routing bridge

To keep representation/control separation explicit:

  • JEPA-like adapter emits unsigned latent deviation and uncertainty streams only.
  • REE computes signed decomposition downstream (for example, harm/benefit directional channels) using claim-owned control-plane rules.
  • Precision routing consumes pe_latent + the confidence channel (derived from calibrated proxy bank inputs such as uncertainty_latent) + context tags; it is not learned implicitly inside the JEPA-like adapter.

Practical interpretation:

  • A JEPA-like reference profile gives the mismatch stream (what was wrong) and can expose dispersion (how many futures looked plausible).
  • REE converts that into a confidence channel (uncertainty-derived precision: how strongly to trust this error for control) with explicit, auditable transforms.

Required Knobs (Config Contract)

The representation adapter must expose, at minimum:

  • latent_dim
  • prediction_horizon
  • context_window
  • action_conditioned (bool)
  • uncertainty_head (bool)
  • update_rate_e1_proxy (slow path)
  • update_rate_e2_proxy (fast path)
  • residual_export_mode (global_only per_mask per_token)
  • uncertainty_estimator (none dispersion ensemble head)

REE-side control knobs remain out-of-scope for JEPA-like reference adapters and must be separately configured in control-plane docs.

Evaluation Contract

Required metrics (stable keys)

  • latent_prediction_error_mean
  • latent_prediction_error_p95
  • latent_rollout_consistency_rate
  • latent_uncertainty_calibration_error (if uncertainty head present)
  • action_conditioned_delta_error (if action-conditioned enabled)
  • latent_residual_coverage_rate (fraction of predictions with exported residual trace)
  • precision_input_completeness_rate (fraction of steps with all required PE/uncertainty fields)

Additional required metrics (JEPA-like control-proxy routing profiles)

  • proxy_bank_coverage_rate (fraction of steps with declared proxy provenance)
  • proxy_confidence_calibration_ece (confidence-channel calibration error from proxy-driven confidence)
  • proxy_residual_correlation_abs (absolute correlation between confidence channel and raw residual magnitude)
  • proxy_ablation_control_delta (control quality delta when proxy bank is disabled vs enabled)

Required checks

  • separation check: E1-proxy updates are slower than E2-proxy updates;
  • no direct commitment check: representation outputs cannot bypass E3;
  • attribution readiness check: outputs contain enough trace context for reafference comparison.
  • uncertainty provenance check: every uncertainty value must declare estimator type (dispersion/ensemble/head);
  • signed-PE boundary check: adapter does not emit control-plane valence labels as if they were representation-native authority labels.
  • proxy provenance check: every active control proxy has proxy_bank declaration fields.
  • confidence/residual separability check: confidence channel is not a direct alias of residual magnitude.
  • ablation utility check: at least one declared proxy improves control/attribution behavior under matched seeds.

Failure Modes and Handling

  • FM-INT-001: Latent shortcut lock-in
    Symptom: low training error with poor transfer under intervention shifts.
    Handling: adversarial environment variants + counterfactual action sweeps.

  • FM-INT-002: Representation adapter overreach into control
    Symptom: adapter performs implicit action selection before E3.
    Handling: strict interface check that adapter emits predictions only.

  • FM-INT-003: Timescale collapse
    Symptom: E1 and E2 update channels become indistinguishable.
    Handling: enforce update-rate ratio thresholds and reject runs failing separation.

  • FM-INT-004: Responsibility-blind predictions
    Symptom: no usable attribution deltas for self-impact loops.
    Handling: require action-conditioned traces and reafference comparison hooks in output pack.

Evidence Routing

Primary claim linkage:

  • MECH-057 (control-completion requirement)
  • Q-012 (falsifiability of control necessity)

Evidence should be recorded via:

  • literature: evidence/literature/targeted_review_mech_057/**
  • experiments: evidence/experiments/claim_evidence.v1.json

Acceptance Gate for Adoption

Adopt a JEPA-like representational adapter profile for E1/E2 interfaces only when all conditions hold:

  • literature support for representation adequacy remains net-positive,
  • experimental runs pass interface checks and stable-key metric checks,
  • no unresolved evidence that the reference adapter bypasses REE commitment/control ownership,
  • governance queue has explicit approval for adoption decision.
  • adoption remains interface/reference scoped and does not make JEPA part of REE architecture.
  • IMPL-022
  • IMPL-021
  • IMPL-020
  • MECH-057
  • Q-012
  • ARC-001
  • ARC-002
  • ARC-005
  • ARC-015

Primary Source Anchors

The following source families were used to define the signal contract:

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