Reafference Comparator Family

Claim: ARC-061 (self_attribution.reafference_comparator_family, candidate). This document is the design-doc backing for that claim.

Describes the structural relationship between three substrate-level implementations of the same reafference-cancellation motif in REE V3. The individual constituent claims (MECH-095, ARC-058, MECH-094) are registered separately; ARC-061 captures their shared architecture.

Last updated: 2026-05-03

The core motif

Von Holst’s reafference principle (1950) names a specific computational problem: a self-monitoring agent must distinguish signals produced by its own actions (reafference) from signals produced by the world acting on it (exafference). The solution is a forward model that predicts the expected self-caused signal; the comparator cancels it; what remains is the world’s contribution. Without this, every self-generated movement is experienced as externally imposed, and every external event is either attributed to self or lost in the noise of action.

This problem does not resolve once at the level of motor commands and stop. It recurs at every level of representational abstraction at which a self-monitoring system operates:

  • Motor/action level: did this body change result from my motor command or from the environment?
  • Interoceptive level: did this affective/somatic state change result from my own internal dynamics or from an external aversive input?
  • Propositional level: did this thought/belief originate from my own generative process or from something injected from outside?

The same motif answers all three:

Forward model F: given current state + agent's action, predict expected signal
Comparator C:    observed signal - F(state, action) = residual
Authorship tag T: if |residual| < threshold -> self-generated
                  if |residual| > threshold -> other-generated

In REE V3, each level is implemented with its own substrate, its own threshold calibration, and its own failure modes — but the architectural pattern is identical.

Three instantiations

Level 1 — Motor / action comparator

Claim: MECH-095 (tpj.agency_detection_comparator, status: active, epistemic_category: substrate_ceiling)

What it computes: E2_harm_s(z, a_actual) - E2_harm_s(z, a_cf) — the counterfactual gap on the harm forward model. Self-caused state change has a distinctive causal signature; environmental change does not. The comparator licenses the authorship tag when that gap exceeds threshold.

REE substrate: E2_harm_s (ARC-033 independent-per-stream path, or ARC-058 shared-trunk variant). The counterfactual branch is the SD-003/SD-029 pipeline.

Biological analogue: Cerebellum (efference-copy cancellation of motor-induced somatosensory change — Blakemore/Frith 2000); SMA/TPJ network for social agency attribution (Pitcher & Ungerleider 2021, third visual pathway / STS for dynamic social perception).

Failure mode (clinical): Schneiderian first-rank passivity phenomena. Shallow comparator slope across S/N → symmetric over-attribution (agent claims responsibility for environmental changes) AND under-attribution (agent fails to register own causal footprint). Same mechanism, opposite error directions depending on noise regime (Asai 2016).

V3 status: substrate_ceiling — CausalGridWorldV3’s “other-caused” change is too thin (single scheduled hazard source) for the comparator to produce a legible signal. V3-EXQ-506 confirmed the ceiling signature (C4-only-PASS, r2_s_to_a ~ 0.99 invariant to action type). SD-047 substrate now IMPLEMENTED 2026-05-03 (calibration ratio 1.95:1 within 1:1-2:1 target band; bit-identical OFF verified; 4-arm sweep validation queued as V3-EXQ-509/510).

Unblocking SD: SD-047 (environment.multi_source_dynamics, IMPLEMENTED 2026-05-03). Design doc: docs/architecture/sd_047_multi_source_dynamics.md.

Level 2 — Interoceptive / affective comparator

Claim: ARC-058 (harm_stream.shared_forward_trunk, status: candidate, v3_pending; competes with ARC-033)

What it computes: Modality-independent unsigned aversive prediction error from a shared HarmForwardTrunk, with stream-specific heads reading out signed per-modality residuals (z_harm_s somatic, z_harm_a affective). The interoceptive comparator tags whether a somatic/affective state change is predicted by current body state and ongoing dynamics (self-generated: injury, drive depletion, pain progression) or anomalous (other-generated: unexpected assault, unpredicted aversive).

REE substrate: ARC-058 HarmForwardTrunk (shared-substrate hypothesis) or ARC-033 independent-per-stream paths (competing). The relevant authorship-tag output feeds z_harm_a stream and modulates drive state attribution.

Biological analogue: Anterior insular cortex (Craig 2009 interoceptive- salience hub); Seth & Friston 2016 active interoceptive inference framework (predictive processing of visceral states with free-energy-minimising body-state regulation). Failure mode: affective prediction errors attributed to external sources when the interoceptive comparator under-predicts self-generated affective change (FND, conversion disorder — Lyndon 2026; negative-symptom self-other indistinctness — Jeganathan & Breakspear 2021).

Failure mode (clinical): Somatic / affective passivity — “made feelings”, made emotions, somatic passivity experiences. The interoceptive stream cannot tag its own affective changes as self-generated; they are experienced as externally imposed.

V3 status: V3-tractable in principle (SD-011 dual-stream substrate is IMPLEMENTED; z_harm_a pathway active). The comparator itself is not yet a discrete mechanism — ARC-058 is the architectural claim that would register it, and is currently competing with ARC-033. V3-EXQ-508 (body-damage ablation) weakened ARC-033 at its borderline boundary, which is evidence in favour of the ARC-058 shared-trunk path.

Gap: No SD claim registered for the interoceptive-comparator substrate enrichment (analogous to SD-047 for Level 1). The Level 2 comparator needs its own substrate-noise calibration: stochastic perturbation of the body-damage stream (interoceptive noise) to probe the same inverted-U calibration curve. This is a pending work item — see section below.

Level 3 — Propositional / simulation tag

Claim: MECH-094 (default_mode.simulation_real_distinction_write_profile, status: stable, confirmed_established)

What it computes: A categorical write-gate φ(z) that marks simulation-mode content with hypothesis_tag = True, preventing it from accumulating in the viability residue field as committed experience. Real-experience content carries hypothesis_tag = False and writes normally. The gate is the authorship tag for propositional/simulation content.

REE substrate: hypothesis_tag boolean threaded through HippocampalModule.replay(), tick() simulation paths, and SD-032e pACC autonomic controller. Tested directly in V3-EXQ-499 (forced- injection discriminative pair: ARM_A with tag active suppresses 50 drive- bias accumulation events, ARM_B without tag accumulates normally).

Biological analogue: PFC/mPFC source monitoring (Corlett 2025 aberrant- salience framework); ACC-hypothesis-tag function (MECH-094 notes); pretend- play tag co-exercise with ARC-049 (MECH-198/199/200/201 maturational cluster). Tag loss = confabulation (MECH-094 primary) and psychosis- adjacent states (MECH-201 notes). Distinct from MECH-095 action-passivity by being content-proximal rather than action-proximal.

Failure mode (clinical): Thought insertion/withdrawal (internally generated thought lacking hypothesis tag, experienced as external); confabulation (replay content writes as real experience). Mechanistically distinct from action-passivity (Level 1) and affective passivity (Level 2).

V3 status: Stable, confirmed_established. No substrate gap — MECH-094 is the most mature member of the family in V3.

Shared diagnostic principles

These principles apply to all three levels. They derive from the common architectural substrate (noise-dependent forward-model comparator) rather than from any level-specific implementation:

1. Symmetric over/under-attribution as calibration indicator

A comparator with a shallow slope produces both over-attribution (agent claims credit for external change) and under-attribution (agent misses own causal footprint) across S/N regimes — not just one direction. The symmetric error pattern is the diagnostic fingerprint of calibration failure, not of a unidirectional bias.

Asai (2016, Psychiatry Research): agency-rating slope is a continuous function of S/N. Schizotypal participants show shallower slopes that produce both error types symmetrically across contexts.

Implication for REE monitoring: an agent whose Level 1 or Level 2 comparator shows systematic error in only one direction probably has a floor or ceiling problem; symmetric errors across conditions indicate a slope (calibration) problem.

2. Non-monotonic noise-dependence (stochastic resonance)

Comparator competence has an optimal substrate-noise level, not a maximum-tolerance level. Too little environmental noise leaves the comparator with no signal to detect; too much noise overwhelms the comparison. An intermediate noise level is the calibration optimum.

Ward et al. (2010, PLoS ONE): 40 Hz transient response in bilateral auditory cortex is enhanced — not degraded — by added broadband noise at an intermediate level. Synchronization between cortical regions shows the same inverted-U.

Implication for SD-047 validation: the 4-arm noise-level sweep (OFF / 0.25x / 1.0x / 4.0x) probes this curve directly. The same sweep protocol should be used for Level 2 calibration when its SD is implemented.

3. Single substrate pathology → bidirectional surface consequences

One substrate-level problem (E/I imbalance, precision miscalibration) produces multiple surface failure modes in opposite directions, because the comparator has multiple consumers that each exhibit the imbalance differently.

Jardri & Deneve (2013, Brain): circular belief propagation from E/I imbalance produces hallucinations AND paradoxical immunity to perceptual illusions from the same mechanism.

Nassar et al. (2021, PNAS): all-or-nothing belief updating produces both over-sensitivity (in volatile environments) and under-sensitivity (in stable environments) from the same computational pathology.

Implication for taxonomy: when an agent shows contradictory attribution errors across contexts, the first diagnostic hypothesis should be a single substrate-calibration problem, not two independent faults.

4. Substrate enrichment precedes content-level intervention

Because the failure is at the substrate-calibration level (comparator slope), it cannot be corrected by content-level interventions (changing what the agent believes, editing priors). SD-047 must land before MECH-095 can be honestly tested — not because the claim is wrong, but because the substrate is too thin to generate the needed signal.

Implication for clinical analogy: substrate-calibration interventions (pharmacological, environmental richness) should improve comparator function before content-level CBT addresses the resulting beliefs (P-018 in psychiatric_predictions.md).

Cross-level summary

  Level 1 (motor/action) Level 2 (interoceptive) Level 3 (propositional)
Claim MECH-095 ARC-058 MECH-094
Subject tpj.agency_detection_comparator harm_stream.shared_forward_trunk simulation_real_distinction
REE substrate E2_harm_s counterfactual gap HarmForwardTrunk + z_harm_a hypothesis_tag write gate
Timescale ms – seconds seconds – minutes episode
Modality sensorimotor somatic/visceral conceptual/propositional
Biological cerebellum/SMA/TPJ insula/AIC PFC/ACC/mPFC
Failure mode (clinical) passivity of action, made impulses made feelings, somatic passivity thought insertion, confabulation
Failure direction symmetric over/under per S/N regime symmetric over/under per noise tag loss -> real attribution of simulation
Authorship tag cf_gap > threshold affective-PE > threshold hypothesis_tag = True/False
V3 status substrate_ceiling, SD-047 IMPLEMENTED V3-tractable, SD-048 pending stable, confirmed_established
Calibration SD SD-047 (implemented 2026-05-03) SD-048 (ready) not applicable
Key literature Asai 2016, Blakemore/Frith 2000 Seth/Friston 2016, Lyndon 2026 Corlett 2025, MECH-094 notes

Implementation gaps

Gap 1: Level 2 substrate enrichment SD (highest priority)

Level 2 has no SD claim. The interoceptive comparator’s substrate-calibration test requires controlled perturbation of the body-damage / affective stream — an interoceptive analog of SD-047’s multi-source noise sweep. Concretely: stochastic variation in z_harm_a background amplitude, independent of the agent’s actions, at calibrated intensity. This exposes whether the Level 2 comparator can separate agent-caused affective change from environmental-origin affective perturbation.

Suggested claim: SD-NNN (body.interoceptive_noise_dynamics), depending on SD-022 (body-damage substrate, IMPLEMENTED) and SD-011 (dual nociceptive streams, IMPLEMENTED). V3-tractable.

Gap 2: Cross-level interaction

V3 can probe each level independently but cannot test whether a failure at one level propagates to another. In biological systems, action-passivity and thought insertion co-occur in schizophrenia — suggesting either shared substrate or cross-level dependency. The question of whether Level 1 calibration failure induces Level 3 tag instability (or vice versa) cannot be answered with independent single-level probes.

This is V4-1 territory (multi-agent ecology + full simultaneous activation of all three comparator levels). See docs/architecture/v4_spec.md.

Gap 3: Cross-level interaction experiment

Level 1 (SD-047) and Level 2 (SD-048) are independently implementable and will be validated separately. Once both land, a joint experiment probing the cross-level interaction becomes the next step: does Level 1 miscalibration propagate to Level 2 tag instability, or are the levels cleanly decoupled? That question cannot be answered with single-level probes. Deferred pending both SDs validated.

V4 implications

When genuine other-agents are present (V4-1), all three comparator levels activate simultaneously and interact:

  • Level 1 must discriminate self-caused from agent-caused from environment-caused change (three-way, not two-way attribution)
  • Level 2 must attribute affective contagion (other-agent’s distress state producing prediction error in own interoceptive stream) as other-generated
  • Level 3 must maintain the simulation/real tag for internally generated models of other agents’ states (theory-of-mind simulations carry hypothesis_tag = True)

The shared diagnostic principles (symmetric over/under-attribution, non-monotonic noise-dependence, single substrate -> bidirectional surface) all generalise to the multi-agent case. In particular, a Level 2 calibration failure in the presence of other-agents produces affective contagion indistinguishable from own affective experience — a computational model of the empathy dysregulation cluster in borderline/dissociative presentations.

That claim cohort is currently V4-bound. The V3 work (SD-047 for Level 1, SD-048 (Level 2 interoceptive noise dynamics)) lays the calibration infrastructure that V4 social probes will need.

  • docs/architecture/sd_047_multi_source_dynamics.md — SD-047 specification (Level 1 substrate enrichment)
  • docs/architecture/substrate_roadmap.md — V3 enrichment planning
  • docs/architecture/v4_spec.md — V4 multi-agent ecology scope
  • docs/architecture/control_plane.md — hypothesis_tag threading (MECH-094 Level 3 implementation)
  • REE_assembly/evidence/literature/targeted_review_connectome_mech_095/ — 15 entries grounding Level 1 substrate-ceiling diagnosis
  • REE_assembly/evidence/literature/targeted_review_arc_058/ — Level 2 interoceptive-comparator literature
  • ai-cognitive-failure-taxonomy/failure_modes/agency_attribution_failure.md — taxonomy entry anchored on Level 1 failure (Asai 2016)
  • ai-cognitive-failure-taxonomy/docs/psychiatric_predictions.md — P-017 / P-018 covering comparator-slope predictions

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