SD-047: Multi-Source Environmental Dynamics

Claim ID: SD-047 Subject: environment.multi_source_dynamics Status: IMPLEMENTED 2026-05-03 Registered: 2026-05-03 Implemented: 2026-05-03 (ree-v3 / ree_core/environment/causal_grid_world.py; substrate landed bit-identical OFF; 4-arm sweep smoke verified ARM_0 legacy-parity, ARM_2 calibration ratio 1.95:1 within target band) Origin: Substrate-roadmap M-priority #4 (docs/architecture/substrate_roadmap.md). Trigger: V3-EXQ-506 (2026-05-03) confirmed the substrate-ceiling signature flagged in MECH-095’s evidence_quality_note (2026-05-02) — the agency-detection comparator PASSes the agent-vs-correlated condition (C4) but fails the agent-vs-env, agent-vs- collateral, and env-vs-correlated conditions (C1-C3). The diagnosis is that CausalGridWorldV3’s “other-caused” change is too thin a causal background for the comparator to learn a robust not-self baseline. Depends on: MECH-095, SD-022

Problem

CausalGridWorldV3’s “other-caused” state change consists of three sources only:

  • Hazard drift — deterministic per-tick step on the scheduled-hazard field.
  • Resource respawn — stochastic timer-driven respawn after consumption.
  • Scheduled hazard injection — fixed per-episode hazard placement.

This is a thin causal background. For agency-detection-comparator claims like MECH-095, the agent’s counterfactual signature E2_harm_s(z_t, a_actual) - E2_harm_s(z_t, a_cf) only separates from environmental noise when there is environmental noise. With sparse, scheduled events, the comparator may match in narrow regimes (V3-EXQ-047k PASS, narrow contact-recall window) but fails in the broader landscape (V3-EXQ-506: agent vs env, agent vs collateral, env vs correlated all collapse). The substrate is too clean — agent action accounts for most of the per-tick change, and the comparator can satisfy C4 (agent-caused vs env-correlated) trivially while still missing the structural distinction the claim requires.

The same thinness limits MECH-098 (reafference cancellation, which is for noisy multi-source backgrounds) and MECH-099 (downstream agency attribution, which has nothing meaningful to attribute when env-causation is sparse and scheduled).

Architectural reading: these are not claim errors; they are substrate- ceiling failures. The claims are V3-tractable in principle but the substrate’s causal vocabulary is too narrow. This is the diagnosis captured by epistemic_category: substrate_ceiling in Phase 3 wave 2.

Mechanism

SD-047 adds three concurrent stochastic event sources to CausalGridWorldV3, each operating at a distinct spatial / temporal scale so the comparator must learn structural agent-vs-env signatures rather than a single filterable noise pattern.

Source 1: Spatial perturbation field (weather analog)

A slowly-evolving 2D scalar field that perturbs per-cell hazard intensity. Implementation: AR(1) noise on a coarse spatial grid (e.g. 4x4 super-cells covering the 8x8 world), updated per tick with autocorrelation alpha_weather (default ~0.95 – slow drift). Perturbation magnitude scaled to sigma_weather so the per-tick change in hazard intensity is small relative to scheduled hazard appearance but persistent across hundreds of ticks.

Signature: continuous-valued, spatially smooth, temporally autocorrelated, agent-independent.

Source 2: Probabilistic transient events (trap analog)

A low-rate Poisson process producing transient per-cell hazard appearances and disappearances independent of agent action. Per-tick per-cell appearance probability p_transient_appear (default ~1e-3 – ~1 transient per ~12 cells per episode at 8x8) and per-tick disappearance probability p_transient_disappear (default ~0.1 – mean lifespan ~10 ticks).

Signature: discrete, spatially-pointwise, temporally short-lived, agent-independent.

Source 3: Background drift sources

A small number n_drift_sources (default 1-2) of independent moving objects following their own dynamics. Each drift source occupies a single cell and steps with its own random-walk / drift policy (drift_policy: random_walk | linear_drift | levy_walk).

Signature: discrete, mobile, temporally autocorrelated, agent- independent.

Why three sources?

A single noise source can be cancelled by a simple temporal or spatial filter. Three sources at different scales force the comparator to learn structural signatures of agent-causation vs env-causation – which is the claim under test. The agent action remains the discrete, focal, agent-controlled signal class; multi-source dynamics provide the not- self baseline against which agency detection is meaningful.

Calibration target

When enabled, env-caused state-change events should occur at roughly comparable rate to agent-caused state changes per episode (target: 1:1 to 2:1 env:agent change events). Below that, the comparator trivially satisfies its criteria; above it, the agent’s own signal is overwhelmed and policy learning collapses. Calibration is empirical (set per-source intensity parameters so an episode of random-policy agent produces ~equal env-caused and agent-caused state-change ticks).

Child mechanisms (deferred)

If SD-047 is implemented and validated, child claims register the distinct signal classes:

MECH-(TBD): spatial-perturbation continuous-noise class

  • Statement: Continuous, smooth, autocorrelated env-caused state change forms a signal class distinct from action-caused state change along smoothness and autocorrelation features.
  • Used by MECH-098 reafference cancellation as the “afference noise” channel.

MECH-(TBD): transient-event discrete-noise class

  • Statement: Discrete, spatially pointwise, short-lived env-caused state change forms a signal class distinct from action-caused state change along sparsity and lifespan features.
  • Used by MECH-095 agency-detection comparator as the principal not-self baseline.

These children are deferred to the implementation phase – registering them now without an implementation invites the same overreach pattern that produced SD-003’s premature elaborate scaffolding (28 FAILs before substrate caught up).

Architecture context

SD-047 sits at the environmental substrate layer, parallel to SD-022 (body-damage substrate), SD-029 (scheduled external hazard), SD-035 (amygdala analog stream substrate). It does not modify the agent or any internal latent: it changes only what the env feeds the agent through obs_dict and the underlying hazard field.

Distinct from V4 multi-agent ecology. Full social / multi-agent substrate (other agents whose causation is genuinely intentional) lives in v4_spec.md (V4-1). SD-047 is the V3-tractable version: env causation that is non-trivial but non-intentional. This is exactly the right level for agency-detection comparator claims, which test the agent vs not- agent distinction; the agent vs other-agent distinction is a separate V4-bound claim.

Distinct from differentiated coping channels (roadmap M-priority #5 / SD candidate environment.differentiated_coping_channels). That SD addresses MECH-102’s substrate-ceiling by giving the agent multiple distinct intervention modalities. SD-047 addresses MECH-095’s substrate-ceiling by giving the env multiple distinct causal sources. Independently implementable.

What this SD enables

Primary unblock:

  • MECH-095 (substrate_ceiling, status: active) — the TPJ agency- detection comparator can be honestly tested. Pre-registered prediction: under SD-047, V3-EXQ-506-equivalent C1-C3 conditions should PASS where they currently FAIL.

Secondary unblocks:

  • MECH-098 (reafference cancellation) — finally has a noisy- background test condition. Currently has no honest substrate.
  • MECH-099 (downstream agency attribution) — has something meaningful to attribute.

Cleaner training distribution:

  • E2_harm_s under SD-013 interventional loss currently sees clean action-controlled vs scheduled-hazard transitions. SD-047 exposes it to mixed-causality transitions where attribution is genuinely ambiguous. The cf_gap_ratio bottleneck identified in V3-EXQ-330 (cf_gap_ratio ~ 1.3-1.5x, target > 2.0x) may partly reflect the too-clean training distribution; SD-047 + SD-013 jointly is a candidate fix.

Indirectly relevant:

  • ARC-033 (sensory-discriminative harm forward model, status: provisional). The recent V3-EXQ-508 borderline FAIL flagged ARC-058 (somatic/affective dissociation per Wager 2013, Hofbauer 2001) as a successor direction. Robust dissociation evaluation benefits from richer multi-source backgrounds.

Validation experiment (deferred)

Pre-registered protocol for SD-047 validation — noise-level sweep:

The original draft proposed a binary ON-vs-OFF discriminative pair. The pre-implementation lit-pull (evidence/literature/targeted_review_sd_047/) revised that protocol on two independent grounds. Both revisions are load-bearing for interpretability.

Revision 1: sweep, not binary — Asai 2016 (DOI: 10.1016/j.psychres.2016.10.082) shows that agency judgment is a non-monotonic function of the signal-to-noise ratio of self-originated contribution to a continuously varying stimulus. Schizotypal participants exhibit a shallower regression slope of agency rating against self-other discriminability and produce both over- AND under-attribution errors symmetrically depending on the S/N regime. The implication for SD-047 is that the comparator’s apparent competence is non-monotonic in env noise level: too little env noise and the comparator never sees the regime where slope is informative (V3-EXQ-506 failure mode); too much and the agent’s own signal is overwhelmed and slope flattens again. A binary ON/OFF test is therefore under-specified — a flat ON-vs-OFF result would be ambiguous between “SD-047 works” and “SD-047 overshot calibration.”

The corrected protocol is a multi-arm noise-level sweep at fixed calibration parameters along all three sources. Suggested levels (to be tuned at implementation):

  • ARM_0 (OFF baseline): multi_source_dynamics_enabled=False. Expected to replicate V3-EXQ-506’s C1-C3 FAIL pattern.
  • ARM_1 (low): all sources at 0.25x default intensity.
  • ARM_2 (mid, default): all sources at 1.0x default intensity (the calibration target: ~1:1-2:1 env:agent change events per ep).
  • ARM_3 (high): all sources at 4.0x default intensity.

Revised pre-registered prediction:

  • ARM_0 C1, C2, C3 FAIL (replication).
  • C1, C2, C3 pass-rate forms an inverted U across ARM_1 → ARM_2 → ARM_3, peaking at or near ARM_2 (default calibration).
  • Asai’s slope analog (cf_gap_event / cf_gap_quiet ratio across the four held-out conditions) is highest at the U’s peak and degrades symmetrically toward both extremes.

Revision 2: explicit Woo/Spelke falsifier branch — Woo, Liu, Spelke 2023 (DOI: 10.1111/desc.13453) showed that 3-month-old prereaching infants attribute goal-directedness to others’ reaching actions from sparse, clean visual stimuli, before they themselves can reach. This challenges the strong reading of substrate-ceiling: if clean stimuli support agency attribution at very young ages, the discriminative features that matter may be motion-contingency and goal-directedness rather than smoothness/sparsity/lifespan. Those features are V4-bound (genuine other-agents) rather than V3-tractable.

This routes a specific failure pattern to a specific architectural diagnosis rather than to “SD-047 is broken.”

Revised interpretation grid:

ARM_2 outcome ARM_0 outcome Reading
C1, C2, C3 PASS C1, C2, C3 FAIL SD-047 validated. Substrate-ceiling diagnosis correct. Lift MECH-095 substrate_ceiling flag; mark MECH-098 evaluable; update evidence_quality_notes.
Inverted U with ARM_2 peak Same as above SD-047 validated with calibration confirmation. Asai non-monotonicity directly observed; default calibration is on the optimum.
ARM_1 or ARM_3 peaks instead of ARM_2 C1-C3 FAIL SD-047 validated, calibration miscalibrated. Recalibrate per-source intensities; do not draw architectural conclusions from this run.
Flat curve, all arms FAIL C1-C3 FAIL Falsification of substrate-ceiling diagnosis as smoothness/autocorrelation/sparsity feature gap. This is the Woo/Spelke branch: the comparator’s load-bearing features are likely motion-contingency / goal-directedness, which V3 substrate cannot deliver without genuine other-agents. Route MECH-095 from substrate_ceiling (V3-tractable) to substrate_conditional with V4 dependency on V4-1 multi-agent ecology. SD-047 retired or kept as ancillary substrate enrichment.
Flat curve, all arms PASS All-PASS Substrate-ceiling diagnosis was wrong but in the opposite direction. Investigate whether V3-EXQ-506 had a non-substrate confound (training pathology, seed pathology, etc.).

This is a falsifiable substrate test in the strong sense: each row of the interpretation grid maps a distinct experimental signature to a distinct architectural conclusion, including a clean route from falsification to a different SD candidate (V4-1 multi-agent ecology). The Woo/Spelke branch is not a graceful loss — it is a productive finding that reroutes substrate work to V4.

Lit-pull provenance: the entries supporting this revised protocol are in evidence/literature/targeted_review_sd_047/: Asai 2016 (load-bearing, supports conf=0.83), Sawtell 2010 (mechanism grounding, supports 0.80), Pitcher & Ungerleider 2021 (pathway specialisation, supports 0.78), Davidson & Wolpert 2005 (forward-model scaffold, supports 0.70), Woo/Liu/Spelke 2023 (falsifier, weakens 0.65). SD-047 lit_conf was 0.0 before this pull and is 0.796 after.

Implementation surface

ree-v3 / ree_core/env/causal_grid_world_v3.py (or equivalent env module):

@dataclass
class WeatherFieldConfig:
    enabled: bool = False
    grid_super_cells: int = 4   # coarse grid resolution
    alpha_ar1: float = 0.95     # autocorrelation
    sigma: float = 0.05         # per-cell perturbation magnitude

@dataclass
class TransientEventConfig:
    enabled: bool = False
    p_appear: float = 1e-3
    p_disappear: float = 0.1
    intensity: float = 1.0

@dataclass
class BackgroundDriftConfig:
    enabled: bool = False
    n_sources: int = 1
    drift_policy: Literal["random_walk", "linear_drift", "levy_walk"] = "random_walk"

@dataclass
class MultiSourceDynamicsConfig:
    enabled: bool = False  # master switch; bit-identical OFF
    weather: WeatherFieldConfig = field(default_factory=WeatherFieldConfig)
    transients: TransientEventConfig = field(default_factory=TransientEventConfig)
    drift: BackgroundDriftConfig = field(default_factory=BackgroundDriftConfig)

Bit-identical OFF requirement: when enabled=False, the env’s internal RNG sequence and per-tick state-update path must match the pre-SD-047 baseline exactly. Add the multi-source RNG draws inside an if cfg.enabled: branch; do not draw RNG values that get discarded. This preserves all existing experiment reproducibility.

Per-source bit-identical OFF: each sub-source can be enabled independently for ablation studies (which source contributes most to unblocking C1/C2/C3).

  • MECH-095 — substrate_ceiling claim this SD primarily unblocks. evidence_quality_note already documents the substrate-ceiling diagnosis (2026-05-02 annotation; 2026-05-03 V3-EXQ-506 confirmation).
  • MECH-098 — reafference cancellation, currently lacks honest substrate.
  • MECH-099 — downstream agency attribution.
  • ARC-033 — provisional, would benefit from richer env for ARC-058 successor direction (Wager 2013, Hofbauer 2001 lit-pull).
  • SD-022 — body-damage substrate (depends_on; SD-047 layers on top of, does not replace).
  • SD-029 — scheduled external hazard (depends_on; SD-047 adds unscheduled env causation alongside).
  • environment.differentiated_coping_channels (substrate-roadmap M-priority #5; not yet registered) — parallel SD addressing MECH-102’s substrate_ceiling. Independently implementable.

SD-047 addresses Level 1 (motor/action comparator) of the reafference comparator family. For the full family context and cross-level diagnostic principles, see: docs/architecture/reafference_comparator_family.md

Lit-pull recommendation (pre-implementation)

Before substrate work begins, commission a focused lit-pull on:

  • agency-detection under multi-source environmental noise (developmental + cognitive psychology on infant agency-perception in noisy environments)
  • reafference cancellation under stochastic afference (cerebellum + parietal cortex)

This follows the “biology before formal definitions” rule (feedback_biology_before_formal_definitions.md): SD-047 is operationally simple but the signature features the comparator should learn (smoothness, autocorrelation, sparsity, lifespan) are empirical questions about how biological agency-detection actually discriminates agent from non-agent. Skipping this risks a philosophy-right / mechanism-wrong implementation that produces a clean PASS on a wrong test.

Anchor candidates:

  • Saxe & Carey 2006, Cog. Dev. (infant agency perception)
  • Blakemore et al. 1998, Nat. Neurosci. (reafference cancellation cerebellar)
  • Pitcher & Ungerleider 2021, TICS (lateral cortex agency network)

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