SD-048: Interoceptive Noise Dynamics

Claim ID: SD-048 Subject: body.interoceptive_noise_dynamics Status: IMPLEMENTED 2026-05-03 (substrate; v3_pending until validation experiment lands) Registered: 2026-05-03 Depends on: SD-011 (dual nociceptive streams, IMPLEMENTED), SD-022 (directional limb damage, IMPLEMENTED) Design doc for: Level 2 of the reafference comparator family (see docs/architecture/reafference_comparator_family.md) Implementation: ree-v3/ree_core/environment/causal_grid_world.py (CausalGridWorldV2 flat __init__ kwargs; readout-side perturbation on harm_obs_a in _get_observation_dict via _apply_interoceptive_noise). See ree-v3/CLAUDE.md SD-048 entry for the full implementation note, config defaults, activation-smoke results, and implementation-choice deviations from this doc.

Problem

SD-022 gave z_harm_a an agent-caused source: limb damage accumulates when the agent moves through hazards and heals slowly. That was the necessary first step — without it, z_harm_a had no agent-caused variance at all (the r2_s_to_a = 0.996 failure documented in SD-022’s functional_restatement).

But the Level 2 comparator problem is the mirror image of the Level 1 problem (MECH-095 / SD-047): the comparator must separate agent-caused body-state change from body-state change that arises independently of what the agent does. After SD-022, all z_harm_a variance is either agent-caused (limb damage from hazard contact) or deterministic (heal_rate = 0.002/step). There is no independent body-state background for the comparator to calibrate against.

The Level 2 claim (ARC-058, HarmForwardTrunk) asserts that the interoceptive forward model can encode an unsigned aversive prediction error shared across body-state sources. That claim is untestable on current substrate: when all interoceptive variance is agent-caused, a trivial forward model E2_harm_a(z, a) = z_harm_a_next passes by memorising the agent’s action-damage coupling, not by learning to separate self-caused from body-noise-caused change.

SD-048 adds three concurrent stochastic body-state perturbation sources that are NOT caused by the agent’s motor output. With these active, the interoceptive comparator must learn to distinguish:

  • Agent-caused: I moved N through a hazard -> limb N is more damaged -> z_harm_a increased
  • Body-generated: autonomic fluctuation / sensitisation spike / fatigue drift -> z_harm_a increased despite no new hazard contact

This is the Level 2 reafference-cancellation problem in its minimal honest form.

Mechanism

SD-048 adds three concurrent stochastic body-state noise sources, each operating at a distinct temporal scale and statistical character. The design mirrors SD-047’s three environmental sources: the comparator must learn structural self/body-noise signatures, not just filter a single noise pattern.

Source 1: Autonomic background fluctuation

Fast, continuous, low-amplitude additive noise on harm_obs_a at every step.

  • Biological analog: beat-to-beat heart rate variability, spontaneous sympathetic fluctuations, background nociceptive sensitisation, the irreducible noise floor of the insular interoceptive map.
  • Statistical character: i.i.d. Gaussian at the step level (no inter-step autocorrelation). Amplitude small relative to agent-caused limb-damage variance.
  • Parameters: autonomic_noise_scale (default 0.02 normalised harm units), applied per step to harm_obs_a directly.
  • Temporal scale: fast (every step).

Source 2: Sensitisation spikes (transient amplification events)

Discrete Poisson-distributed events that transiently amplify harm_obs_a by a multiplicative factor, then decay exponentially.

  • Biological analog: inflammatory sensitisation flares, allodynic episodes, spontaneous visceral pain amplification, central sensitisation bursts. These are not caused by new tissue damage — they are changes in the gain of the interoceptive reporting pathway.
  • Statistical character: Poisson onset (rate sensitisation_rate, default ~0.008/step), multiplicative amplitude (sensitisation_magnitude, default 1.8x), exponential decay with half-life sensitisation_halflife (default 15 steps). Multiple events can overlap (additive amplification).
  • Parameters: sensitisation_rate, sensitisation_magnitude, sensitisation_halflife.
  • Temporal scale: medium (onset discrete, decay ~15-30 steps).

Source 3: Fatigue drift

Slow AR(1) accumulation in a latent background fatigue state that adds to harm_obs_a continuously.

  • Biological analog: metabolic fatigue, allostatic load accumulation, sleep pressure, prolonged sympathetic activation. Fatigue is not the same as tissue damage — it is a slowly-rising body-state cost that persists independently of what the agent is currently doing.
  • Statistical character: AR(1) process with coefficient fatigue_ar_coeff (default 0.995, very slow decay), driven by low-amplitude Gaussian innovations. Contribution to harm_obs_a is additive and proportional to current fatigue state.
  • Parameters: fatigue_ar_coeff, fatigue_noise_scale (default 0.005/step).
  • Temporal scale: slow (episode-length drift, resets at episode end).

Why three sources?

Same logic as SD-047. A single noise source can be filtered by matching its specific statistical signature. Three sources with qualitatively different characters (fast i.i.d. / discrete transient / slow drift) force the comparator to learn a structural signature of agent-causation rather than a specific noise pattern. The sources correspond to three biological timescales:

  • Autonomic (seconds) → step-level jitter
  • Inflammatory/sensitisation (minutes) → episode-phase transient
  • Metabolic/allostatic (hours) → full-episode drift

Calibration target

1:1 to 2:1 body-noise-caused harm-state-change events to agent-caused harm-state-change events per episode, at default calibration (ARM_2).

This mirrors SD-047’s 1:1-2:1 calibration target. The goal is that the background noise is large enough to be non-trivial for the comparator but not so large that agent-caused signal is drowned. Events are defined as steps where |delta_harm_obs_a| > threshold_minor — both agent-caused (hazard contact) and body-noise-caused (sensitisation spike, fatigue step) count.

Child mechanisms (deferred)

These sub-mechanisms will register when the SD implements and when their distinct functional signatures become measurable:

MECH-(TBD): Autonomic-noise interoceptive self-attribution class

The fast Gaussian source contributes a continuous self/body-noise calibration signal to the comparator. Whether ARC-058’s trunk represents the autonomic noise floor as a distinct unsigned-PE class (separate from sensitisation and fatigue) is an empirical question.

MECH-(TBD): Sensitisation-event interoceptive self-attribution class

The discrete Poisson source tests whether the comparator can tag sensitisation-origin harm-obs changes as non-agent-caused. Failure produces “made feelings” phenomenology (body changes attributed to external agency). The sensitisation profile — transient, large amplitude, multiplicative — is structurally distinct from the fatigue profile.

MECH-(TBD): Fatigue-drift interoceptive trajectory class

The slow AR(1) source tests whether the comparator represents fatigue as a self-generated body trajectory (expected given prior fatigue state) or misattributes late-episode fatigue accumulation to environmental harm.

Register these as MECH claims when SD-048 is implemented and the validation experiment runs.

Architecture context

SD-048 sits at the body-state substrate layer, parallel to SD-022 (which added agent-caused body-state variance) and SD-011 (which split the harm stream into z_harm_s and z_harm_a). SD-048 adds the other-caused side of the z_harm_a variance that SD-022 did not address.

Relationship to SD-047: the two SDs address the same general problem (substrate too thin for comparator calibration) at different levels:

  • SD-047: external environment → Level 1 motor/action comparator (MECH-095)
  • SD-048: internal body state → Level 2 interoceptive comparator (ARC-058)

They are independently implementable and can run simultaneously. An experiment with both enabled would test the interaction: does a richer body-state background affect the Level 1 comparator, or are the two levels cleanly decoupled? That cross-level question is not the primary target of either SD individually.

Relationship to ARC-058 vs ARC-033 arbitration: SD-048 is necessary for both paths. Whether the interoceptive comparator uses a shared trunk (ARC-058) or independent per-stream models (ARC-033) is an architectural question that SD-048 does not settle — but neither can be honestly tested until the substrate has independent body-state variance to discriminate.

What this SD enables

  • ARC-058 (harm_stream.shared_forward_trunk, primary) — the HarmForwardTrunk claim can now be tested at substrate level. C1-C3 criteria analogous to MECH-095 (body-noise-event selectivity, self-vs-body-noise gap, interoceptive attributor calibration) become measurable. ARC-058 moves from V3-tractable/no-substrate to directly testable.
  • ARC-033 (competing claim) — same substrate enrichment benefits independent-per-stream comparator testing. The arbitration between ARC-033 and ARC-058 is best run on SD-048-enriched substrate where both can express their architectural signatures.
  • ARC-061 (self_attribution.reafference_comparator_family) — Level 2 contribution. ARC-061 cannot promote until Level 2 has experimental support; SD-048 is the prerequisite.
  • SD-011 sub-claims — several z_harm_a dynamics claims become testable with non-trivial body-state variance.
  • Possibly MECH-112 (wanting/liking dissociation) if fatigue drift interacts differentially with the wanting vs liking gradient fields.

Validation experiment (deferred)

Pre-registered protocol for SD-048 validation — noise-level sweep:

The same architectural logic that motivated SD-047’s 4-arm noise sweep applies here (Asai 2016 non-monotonic comparator competence; Ward 2010 stochastic resonance in human interoceptive EEG). Binary ON/OFF is ambiguous between “SD-048 works” and “SD-048 overshot calibration.”

Primary metric: ARC-058 (or ARC-033 as baseline) z_harm_a comparator performance:

  • C1: Body-noise-event selectivity — comparator output discriminates body-noise-caused z_harm_a spikes from agent-caused spikes (AUC or ROC above chance).
  • C2: Self-vs-body-noise gap — forward model residual is larger for body-noise events than for agent-caused events in matched-amplitude conditions.
  • C3: Interoceptive attributor calibration — comparator output correlates with actual event origin (agent-caused vs body-noise-caused) across conditions.
  • C4: Base z_harm_a forward model accuracy (r2) — must remain high in OFF arm (sanity check that SD-048 doesn’t degrade base model when disabled).

Arms:

  • ARM_0 (OFF baseline): interoceptive_noise_enabled=False. SD-022 only. Should reproduce current substrate ceiling: C1-C3 fail or are untestable (no body-noise events exist), C4 holds.
  • ARM_1 (0.25x): All three sources at 25% of default scale. Below-calibration prediction: C1-C3 may show modest signal but comparator competence is impaired by insufficient background variance.
  • ARM_2 (1.0x default): All sources at default calibration. Predicted C1-C3 PASS. This is the hypothesised optimum from the Ward/Asai non-monotonicity principle.
  • ARM_3 (4.0x): All sources at 4x scale. Above-calibration prediction: body noise drowns agent-caused signal, comparator competence degrades (C1-C3 drift back toward FAIL).

Interpretation grid:

ARM_2 (default) ARM_0 (OFF) Interpretation
C1, C2, C3 PASS C1, C2, C3 FAIL SD-048 validated. Route ARC-058 to standard testable; move Level 2 from pending to evaluable in ARC-061.
Inverted-U with ARM_2 peak Same as above SD-048 validated with calibration confirmation. Non-monotonicity directly observed.
ARM_1 or ARM_3 peaks instead of ARM_2 C1-C3 FAIL SD-048 validated, calibration off. Recalibrate per-source scales; architectural conclusions hold.
Flat curve, all arms FAIL C1-C3 FAIL Falsification of V3-tractable assumption. The interoceptive comparator’s load-bearing features require a richer body-state substrate (temperature, proprioception, nausea, visceral specificity) that CausalGridWorldV3 cannot provide. Route ARC-058 from V3-tractable to substrate_conditional with dependency on a richer body-state substrate. SD-048 retired or kept as ancillary enrichment for SD-011 downstream claims.
ARM_2 PASS, ARM_0 also PASS (trivial comparator) SD-048 irrelevant. Comparator passes without noise, so it was never substrate-gated — revisit C1/C2/C3 threshold calibration and confirm they are non-trivial.

Implementation surface 

# New config fields in REEConfig (ree_core/config.py)
interoceptive_noise_enabled: bool = False           # master switch
interoceptive_noise_scale: float = 1.0              # global multiplier (the arm sweep lever)

# Source 1: autonomic background
autonomic_noise_scale: float = 0.02                 # per-step Gaussian SD (normalised harm units)
autonomic_noise_enabled: bool = True                # per-source ablation switch

# Source 2: sensitisation spikes
sensitisation_rate: float = 0.008                   # Poisson events per step
sensitisation_magnitude: float = 1.8               # multiplicative amplifier
sensitisation_halflife: int = 15                    # steps to half-amplitude decay
sensitisation_enabled: bool = True

# Source 3: fatigue drift
fatigue_ar_coeff: float = 0.995                     # AR(1) persistence
fatigue_noise_scale: float = 0.005                  # per-step innovation SD
fatigue_contribution_weight: float = 0.15           # weight on harm_obs_a additive sum
fatigue_enabled: bool = True

Bit-identical OFF requirement: when interoceptive_noise_enabled=False, the body-state update path must be identical to pre-SD-048 baseline (no RNG draws, no state variables). All three sources must also be individually disableable (autonomic_noise_enabled, etc.) for ablation experiments.

Location: ree_core/environments/causal_grid_world.py, body-state update section (analogous to SD-022’s limb_damage accumulation block). New internal state variables: _fatigue_state: float (reset to 0.0 at episode start), _sensitisation_state: float (active amplification).

Episode reset: _fatigue_state and _sensitisation_state reset to 0.0 at reset(). Sensitisation events are i.i.d. per episode.

  • ARC-058 — primary target. SD-048 is the substrate condition for honest ARC-058 testing.
  • ARC-033 — competing claim. Benefits from same substrate enrichment; arbitration runs on SD-048-enabled substrate.
  • ARC-061 — Level 2 contribution to the reafference comparator family. See docs/architecture/reafference_comparator_family.md.
  • SD-011 — prerequisite. The z_harm_a stream must exist before its interoceptive noise dynamics can be added.
  • SD-022 — prerequisite. Agent-caused body-state variance must exist before other-caused noise is meaningful as a calibration background.
  • SD-047 — Level 1 counterpart. Same architectural logic at the environmental layer. Independently implementable; can run in parallel. Cross-level interaction (joint SD-047 + SD-048) deferred to a later experiment.

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