SD-011: Dual Nociceptive Streams

Claim ID: SD-011 Subject: harm_stream.dual_nociceptive_streams Status: IMPLEMENTED (base 2026-03-30, second source 2026-04-08) Registered: 2026-03-24 Depends on: SD-010, ARC-027 Blocks: SD-003 redesign (EXQ-093/094 confirmed infeasibility of z_world → z_harm bridge)

Problem

SD-010 introduced a single z_harm stream (harm proximity → HarmEncoder → z_harm). This resolves the z_world contamination problem but reveals a second-order problem: a single z_harm stream conflates two neurobiologically incommensurable nociceptive pathways (Melzack & Casey 1968; Craig 2002/2009).

Neurobiological Dissociation

The two-pathway model is experimentally confirmed by Rainville et al. (1997, Science) — the gold-standard dissociation study. Hypnotic modulation of unpleasantness selectively modulates ACC (affective-motivational pathway) without affecting S1 (sensory-discriminative). The two pathways produce different neural signatures, different clinical presentations, and different behavioral functions.

Property z_harm_s (sensory-discriminative) z_harm_a (affective-motivational)
Biological analog Lateral spinothalamic tract → VPL → S1/S2 Medial pathway → CM/PF → ACC/insula/amygdala
Fiber type A-delta (fast, sharp, localized) C-fiber (slow, diffuse, persistent)
Signal content Immediate proximity/intensity Accumulated homeostatic deviation / unpleasantness
Temporal dynamics Fast, discriminative Slow, integrative (EMA tau ≈ 10–30 steps)
Forward-predictable? YES — moving away from hazard reduces proximity NO — accumulated deviation does not reverse quickly
Role in SD-003 Counterfactual attribution target NOT used in counterfactuals
Role in E3 Input to harm attribution Input to commit urgency / motivational gating

Why EXQ-093/094 Found bridge_r2=0

SD-010 places z_world ⊥ z_harm by design — the harm stream is explicitly separated from world processing. Consequently, HarmBridge(z_world → z_harm) is architecturally infeasible: z_world has no information about harm proximity, by construction.

EXQ-093 FAIL (bridge_r2=0) and EXQ-094 FAIL (100x regression) confirmed this. This is not a bug and not fixable by architecture tuning. The SD-003 counterfactual must be redesigned to operate within the harm stream, not across the z_world/z_harm boundary.

Solution

Split z_harm into two dedicated streams:

harm_obs_s (immediate proximity) → HarmEncoderS → z_harm_s   [sensory-discriminative]
harm_obs_a (EMA accumulator)     → HarmEncoderA → z_harm_a   [affective-motivational]

Required Implementation Changes

(a) CausalGridWorldV2 — add harm_obs_a:

# Alongside existing harm_obs (renamed harm_obs_s):
harm_obs_a = EMA(harm_obs_s, tau=10-30_steps)   # accumulated homeostatic deviation

(b) HarmEncoderS + HarmEncoderA:

z_harm_s = HarmEncoderS(harm_obs_s)   # rename current HarmEncoder
z_harm_a = HarmEncoderA(harm_obs_a)   # new; trains on accumulated harm signal

(c) LatentState — add z_harm_a field:

@dataclass
class LatentState:
    z_self:   Tensor   # proprioceptive / interoceptive
    z_world:  Tensor   # exteroceptive world model
    z_harm_s: Tensor   # sensory-discriminative harm (SD-010/011)
    z_harm_a: Tensor   # affective-motivational harm (SD-011)
    z_beta:   Tensor   # affective latent (shared)
    ...

(d) E2_harm_s — new forward model (ARC-033):

class E2HarmS:
    def forward(self, z_harm_s: Tensor, a: Tensor) -> Tensor:
        """Predict z_harm_s at next step: moving away from hazard reduces proximity."""

This is the correct SD-003 counterfactual mechanism. Architecture mirrors E2.world_forward.

(e) E3Selector — separate inputs:

# Attribution (SD-003 redesign):
z_harm_s_actual = E2_harm_s(z_harm_s, a_actual)
z_harm_s_cf     = E2_harm_s(z_harm_s, a_cf)
causal_sig = harm_eval(z_harm_s_actual) - harm_eval(z_harm_s_cf)

# Commit gating (ARC-016 / motivational urgency):
urgency = E3.urgency_from_affective(z_harm_a)   # accumulated threat state scales commit threshold

SD-003 Redesign After SD-011

The SD-003 counterfactual pipeline validated at EXQ-030b (world_forward_r2=0.947) confirmed the counterfactual architecture is sound. It must now be applied to z_harm_s specifically:

# OLD (pre SD-010): operated on z_world -- no longer valid
z_world_cf  = E2.world_forward(z_world, a_cf)
causal_sig  = E3.harm_eval(z_world_actual) - E3.harm_eval(z_world_cf)

# NEW (post SD-011): operates on z_harm_s
z_harm_s_cf = E2_harm_s(z_harm_s, a_cf)
causal_sig  = E3.harm_eval(z_harm_s_actual) - E3.harm_eval(z_harm_s_cf)

Critical constraint: Do NOT attempt HarmBridge(z_world → z_harm_s) counterfactuals. bridge_r2=0 is architectural (z_world ⊥ z_harm by SD-010 design). EXQ-093/094 confirmed this.

What SD-011 Enables

  • SD-003 counterfactual: E2_harm_s provides a learnable forward model on z_harm_s; the counterfactual pipeline is now complete within the harm stream
  • ARC-016 dynamic precision via z_harm_a: accumulated threat state (z_harm_a variance) provides a biologically grounded signal for commit threshold scaling — more precisely than the previous z_world-derived precision
  • MECH-112 wanting/liking balance: z_harm_a as affective urgency stream pairs with z_goal (wanting/approach) for D1/D2 balance in ARC-030
  • INV-007 (language cannot override harm): z_harm_a as a separate persistent channel that is not subject to linguistic modulation or reafference correction

Experiments Blocked by SD-011 Absence

The following experiments have FAILed primarily because the single z_harm stream is insufficient for the SD-003 counterfactual redesign or for separating attribution from motivational gating: EXQ-093, EXQ-094, and approximately 10 additional EXQ numbers pending SD-011 implementation.

These will be re-queued after SD-011 implementation with appropriate "supersedes" fields.

  • SD-011 — this design decision
  • SD-010 — prerequisite (single harm stream that SD-011 splits)
  • ARC-027 — harm stream as dedicated sensory pathway
  • ARC-033 — E2_harm_s forward model for SD-003 counterfactual
  • SD-003 — self-attribution counterfactual (redesign target)
  • ARC-016 — dynamic precision (benefits from z_harm_a as urgency signal)
  • ARC-030 — D1/D2 approach-avoidance balance (z_harm_a as avoidance channel)
  • MECH-112 — wanting/liking dissociation (z_harm_a as affective counterpart)

References

  • Melzack & Casey (1968): Gate control theory — discriminative vs affective pain
  • Craig (2002, 2009): Interoceptive cortex, affective-motivational pain in ACC/insula
  • Rainville et al. (1997, Science): Hypnotic modulation of unpleasantness modulates ACC, not S1 — gold-standard dissociation between the two pathways

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