SD-010: Harm Stream Separation

Claim ID: SD-010 Subject: harm_stream.nociceptive_separation Status: Implemented Registered: 2026-03-19 (after V2 governance cycle) Implemented: 2026-03-20 (EXQ-056c/058b/059c PASS) Depends on: SD-005, SD-007, ARC-027

Problem

After SD-005 introduced the z_self/z_world split, E3 was still evaluating harm via z_world. This caused two inseparable problems:

  1. Residue contamination: ResidueField accumulates world_delta to track causal footprint. If z_world also encodes harm proximity, the residue field becomes a harm-proximity map rather than a causal-attribution map. These are different things — the agent may cause no harm in a hazard-dense environment, and may cause significant harm in a sparse one.

  2. Counterfactual collapse: SD-003 self-attribution requires E2.world_forward(z_world, a_cf) to predict how world-state changes under a counterfactual action. This only works if z_world tracks world structure, not harm. E2 cannot simultaneously predict world-state dynamics AND harm proximity from a single z_world signal.

  3. Reafference interference: SD-007 applies a perspective-correction to z_world (subtract agent-caused optical-flow change). Harm proximity is agent-relative by definition — applying reafference correction to a harm signal would incorrectly subtract the agent’s approach to a hazard from the harm signal, reducing the harm estimate as the agent gets closer. This is backwards.

Root cause cluster: EXQ-027b (reafference over-correction), EXQ-043/044 (calibration collapse), EXQ-047 (SD-005 calibration shortfall) all traced back to fused z_world carrying harm signal.

Solution

Dedicated harm sensory pathway, independent of z_world processing:

# CausalGridWorldV2 observation dict
{
    "body_state":   [...],   # proprioceptive -> z_self encoder
    "world_state":  [...],   # exteroceptive -> z_world encoder (NO harm included)
    "harm_obs":     [...]    # harm proximity -> HarmEncoder -> z_harm (SD-010)
}

# LatentStack.encode() produces:
LatentState(
    z_self   = self_encoder(body_state),
    z_world  = world_encoder(world_state),   # clean; no harm contamination
    z_harm   = harm_encoder(harm_obs),        # SD-010: dedicated harm stream
    ...
)

HarmEncoder is a lightweight MLP trained with direct supervision on proximity labels. E3.harm_eval(z_harm) takes z_harm as primary input — NOT z_world. Reafference correction (SD-007) applies only to z_world; z_harm is explicitly excluded.

Implementation Notes

CausalGridWorldV2 (ree_core/environment/causal_grid_world.py):

  • Emits harm_obs channel alongside world_state
  • harm_obs = proximity-weighted hazard density in agent field of view
  • Ground truth label: transition_type ∈ {agent_caused_hazard, env_caused_hazard, resource, none} (preserved from V2; still used for SD-003 attribution evaluation)

HarmEncoder (ree_core/latent/stack.py):

  • HarmEncoder(harm_obs_dim → harm_dim)
  • Trained on proximity labels (supervised, not reconstruction)
  • No reafference correction applied

E3Selector (ree_core/predictors/e3_selector.py):

  • harm_eval(z_harm: Tensor) → Tensor — primary harm assessment method
  • Dynamic precision: precision = running_mean(E3_prediction_error_variance) over z_harm
  • Commit threshold derived from precision (ARC-016)

What SD-010 Enables

  • Clean z_world for residue: ResidueField now accumulates world_delta (actual causal footprint) without harm-proximity contamination.
  • SD-003 counterfactual on world stream: E2.world_forward(z_world, a_cf) is now interpretable — it predicts how world structure changes, not how harm changes.
  • ARC-016 dynamic precision: E3 prediction error variance over z_harm is a meaningful signal for commit threshold adjustment.
  • Precondition for SD-011: SD-010 defines the single z_harm stream. SD-011 splits it into z_harm_s (sensory-discriminative) + z_harm_a (affective-motivational).

Evidence

Experiment Outcome Criteria met Notes
EXQ-058b PASS 5/5 ARC-027 foundational validation. Pearson r_z_harm high.
EXQ-056c PASS 4/5 SD-010 reafference isolation retest. r_z_harm=0.914. Clean separation confirmed.
EXQ-059c PASS 4/5 MECH-102 via SD-010. contact_rate_reduction ~10x. Approach-gradient avoidance confirmed.

Prior experiments (EXQ-027b, EXQ-043, EXQ-044, EXQ-047) FAILed and traced back to fused z_world as root cause. SD-010 resolved all three FAIL clusters.

What SD-010 Does NOT Solve

SD-010 implements a single z_harm stream. This single stream conflates two neurobiologically incommensurable signals:

  • Sensory-discriminative harm (immediate proximity/intensity — forward-predictable)
  • Affective-motivational harm (accumulated homeostatic deviation — integrative, NOT predictable)

Consequence: the SD-003 counterfactual redesign requires SD-011 (dual nociceptive streams) because HarmBridge(z_world → z_harm) has bridge_r2=0 (z_world ⊥ z_harm by SD-010 design — architectural impossibility confirmed by EXQ-093/094). See sd_011_dual_nociceptive_streams.md.

  • SD-010 — this design decision
  • ARC-027 — harm stream as dedicated sensory pathway (architectural requirement)
  • SD-011 — dual nociceptive stream extension (PENDING)
  • ARC-033 — SD-003 counterfactual requires E2_harm_s forward model (PENDING)
  • MECH-102 — approach-gradient harm avoidance (validated via SD-010)
  • ARC-016 — dynamic precision from E3 prediction error
  • SD-005 — z_self/z_world split (prerequisite)
  • SD-007 — reafference correction (co-designed; does NOT apply to z_harm)

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