Self-Attribution Per Reafferent Stream

Registered: 2026-04-18 Status: Active roadmap document. Supersedes the narrow SD-003 counterfactual architecture.

Summary

Self-attribution in REE is implemented by a single-pass forward-model comparator applied independently to each reafferent latent stream. One general mechanism (MECH-256) over three stream-specific substrate instantiations (SD-029 active, SD-030 / SD-031 V4-deferred). The same forward model per stream is read in two modes (MECH-257): retrospective comparator (attribution) and prospective rollout-scoring (evaluation), arbitrated by a controller signal.

This structure is a deliberate restructuring of the retired two-pass counterfactual architecture (SD-003, 28 FAILs), grounded in a three-pull literature synthesis (2026-04-18, 14 entries across comparator / evaluator / mode-distinction).

Claim Topology

Claim Scope Status Phase Doc
MECH-256 General comparator mechanism, stream-agnostic candidate V3 this doc
MECH-257 Dual-function (comparator vs evaluator), gated readout candidate V3 this doc
SD-029 z_harm_s comparator (active V3 implementation) candidate V3 arc_033_e2_harm_s_forward_model.md
SD-030 z_self comparator (motor-proprioceptive) candidate V4 sd_030_e2_self_forward_model.md
SD-031 z_world comparator (causal footprint) candidate V4 sd_031_e2_world_forward_model.md

SD-003 is superseded by MECH-256 + SD-029 as of 2026-04-18.

The General Mechanism (MECH-256)

For any reafferent latent stream z_x with an efference-copy-driven forward model E2_x:

predicted_z_x = E2_x(z_x_{t-1}, a_actual)
residual_z_x  = z_x_observed - predicted_z_x
causal_sig    = f(residual_z_x, context)

One forward pass. No counterfactual branch. No E2(a_cf) call. The forward prediction on the actually-taken action is the reference; the residual is the agency signal.

Self-caused events produce attenuated residuals (the prediction succeeds); externally-caused events produce large residuals (the prediction fails). Shergill 2003 measures this attenuation at roughly 50% on the force-matching task – partial, not binary.

Per-Stream Substrates

The computational mechanism is uniform across streams; the substrates are stream-specific, matching the anatomical organisation of biological agency attribution:

Stream Substrate Biology
z_harm_s E2_harm_s ACC / insula (nociceptive self-vs-other attribution); descending pain modulation
z_self E2_self (V4) Cerebellum (Wolpert/Miall/Ito); S1/S2 cancellation (Blakemore tickle)
z_world E2_world (V4) Parietal outcome prediction; colliculus / MD-thalamus corollary discharge

Other biological streams (oculomotor, vocal, electrosensory) share the same computational template but are not part of the current REE agent’s sensorimotor envelope.

Dual-Function Reading (MECH-257)

Each E2_x is a single substrate read in two modes:

Mode 1 — Retrospective comparator (attribution, post-action):

residual = z_x_observed - E2_x(z_x_{t-1}, a_actual)

Mode 2 — Prospective rollout-scoring (evaluation, pre-action):

for each candidate action a_i:
    predicted_trajectory = E2_x-rollout(z_x_t, a_i, horizon=k)
    value(a_i) = f(predicted_trajectory)

Same learned weights. A controller signal arbitrates mode. V3 candidate for the gating signal: commitment boundary / MECH-094 hypothesis tag. V4 candidates: dACC EVC (Shenhav 2013/2016) or heartbeat-phase gating (ARC-023).

Biological precedent is strongest in the hippocampus: Diba & Buzsaki 2007 and Dragoi & Tonegawa 2011 show one CA1 population expressing forward (preplay) and reverse (replay) sequences within a single session, phase- governed. Mattar & Daw 2018 gives a single prioritized-memory-access algorithm that produces both forward planning and reverse credit assignment traversals over a shared forward model.

V3 Implementation Target (SD-029)

PASS criteria (event-conditioned, not mean-over-all-steps):

  • C1: forward_r2(E2_harm_s) >= 0.9 on the reafferent z_harm_s stream.
  • C2: Residual magnitude partially attenuated for self-caused vs externally-caused harm events (Shergill partial-attenuation pattern, not a binary gate).
  • C3: Residual signal-to-noise above threshold on approach-to-harm events specifically. Event-conditioned, not mean over all steps – this was the lesson from EXQ-431, where mean cf_gap was uninformative because agent-caused hazard events were sample-starved (counts 1 / 156 / 2 across seeds 42 / 7 / 13).

Training requirement (SD-013): E2_harm_s must see counterfactual perturbations during training to learn P(z_harm_s | do(a)) rather than the correlational P(z_harm_s | z_t, a). Pure observational training produces a biased causal signal in confounded states.

Why SD-003 Was Superseded

Three accumulating reasons led to the supersession decision:

  1. Empirical: 28 FAILs accumulated across V2+V3 iterations with no PASS on the two-pass counterfactual architecture since the z_world → z_harm_s stream migration. EXQ-030b’s original PASS was on the (now-retired) z_world pipeline.

  2. Biological: The three-pull 2026-04-18 synthesis (Frith 2000, Shergill 2003, Blakemore 1998, Haggard 2017) directly evidences a single-pass comparator; the two-pass counterfactual extension has no direct biological precedent. The earlier evidence_quality_note on SD-003 had already flagged this as a caveat; the supersession accepts the caveat as load-bearing.

  3. Architectural: Mean-cf-gap metrics drowned the rare high-signal approach-to-harm states; the single-pass comparator’s signal structure is event-shaped rather than mean-shaped, which aligns with the event-density controls that EXQ-431 showed were required.

The ARC-033 E2_harm_s substrate is carried forward unchanged; it is the concrete forward model SD-029 runs on. SD-003’s evidence chain (EXQ-030b, 115, 166a, 195, 353, 431) remains in the historical record but does not accrue to SD-029 without re-analysis.

First Test for SD-029

Event-conditioned reanalysis of EXQ-330a traces against approach-to-harm events. If the existing traces include sufficient reafference + action data, no new run is required; the analysis is a manifest post-processor. If the sample density is insufficient, a short re-run with event-density controls (guaranteed minimum agent-caused-hazard events per seed) is queued.

Curriculum Substrate (2026-04-21)

The balanced-hazard-event curriculum required for C3/C4 is now implemented as an opt-in env augmentation in CausalGridWorldV2: scheduled_external_hazard_enabled=True with _interval, _prob, _adjacent_only sub-knobs. Disabled by default so all legacy experiments run unchanged. See ree-v3/CLAUDE.md section “SD-029: Balanced Hazard-Event Curriculum (2026-04-21)” for the full data flow. The first validation experiment is V3-EXQ-470.

Status: SD-029 substrate IMPLEMENTED 2026-04-21 – the forward model (ARC-033) and interventional training (SD-013) were already in place; the curriculum gap was the last substrate blocker. Remaining work is experimental (validate C3/C4 under the new curriculum).

Open Questions

  1. Dual-function falsifiability test: does training E2_x jointly for comparator + evaluator degrade mode-specific performance relative to mode-split baselines? Requires an ablation experiment not yet queued.
  2. Nociceptive transfer: the canonical comparator literature is on sensorimotor/tactile/force streams. Extension to nociceptive streams is plausible (PAG/RVM descending modulation shares the efference-copy structure) but not directly demonstrated. This is the main mapping risk for SD-029.
  3. Arbitrator identity: what biological substrate houses the controller signal that gates comparator vs evaluator mode? dACC (Shenhav EVC) and heartbeat phase (ARC-023) are the two V3/V4 candidates.

References

Full literature entries:

  • evidence/literature/targeted_review_sd003_successor_comparator/entries/ (4)
  • evidence/literature/targeted_review_sd003_successor_evaluator/entries/ (5)
  • evidence/literature/targeted_review_e2_dual_function_mode_distinction/entries/ (5)

Legacy architecture (retained for diff):

  • docs/architecture/sd_003_experiment_design.md – the retired two-pass design.

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