SD-021: Descending Pain Modulation — Commitment-Gated z_harm_s Attenuation

Status: candidate
Implementation phase: v3
Implemented: 2026-04-10
Depends on: SD-011, SD-020, ARC-016, MECH-090
Unblocks: MECH-220 (pain-as-control-signal validation)

Claim

When E3 is committed to a trajectory that passes through expected harm (beta_gate elevated, MECH-090), the sensory-discriminative harm stream z_harm_s is precision-downweighted by an attenuation factor descending_attenuation_factor. The affective stream z_harm_a is NOT attenuated — affective load persists through commitment.

This implements the biological descending inhibitory pain modulation pathway: pgACC → PAG → RVM → dorsal horn. The pathway provides top-down suppression of nociceptive input during volitional action through expected harm. The agent commits to a trajectory because it has already evaluated the harm; re-weighting the sensory signal during execution prevents the harm from derailing an already-chosen plan.

Biological Basis

Chen (2023, Front Neural Circuits) documents the descending pain modulatory system as a precision-weighted control pathway, not a simple suppression gate. The key features:

Selectivity: Only z_harm_s (sensory-discriminative, A-delta analog) is attenuated. z_harm_a (affective-motivational, C-fiber analog) persists. This maps to the clinical observation that committed athletes feel pain but are not derailed by it — the motivational urgency signal (z_harm_a) persists but the sensory discrimination signal (z_harm_s) is gated down.
Commitment-dependence: The pathway activates during volitional action through known hazards, not during passive harm exposure. MECH-090 BetaGate elevation is the REE correlate of the committed-action state.
Precision mechanism: The gate does not eliminate z_harm_s; it reduces its effective precision weight. In REE terms this is implemented as a scalar multiplication: z_harm_s_gated = z_harm_s * descending_attenuation_factor where 0 < descending_attenuation_factor < 1.

Architecture

Data flow

env.harm_obs
    |
    v
HarmEncoder --> z_harm_s (LatentState)
                    |
                    v [if beta_gate.is_elevated AND harm_descending_mod_enabled]
                z_harm_s * descending_attenuation_factor
                    |
                    v
              E3 harm_eval, E2_harm_s forward model

z_harm_a is unaffected — it flows from AffectiveHarmEncoder without attenuation.

Config parameters (REEConfig)

harm_descending_mod_enabled: bool = False        # master switch (default: off)
descending_attenuation_factor: float = 0.5       # 0=full suppression, 1=no suppression

Implementation location

ree-v3/ree_core/agent.py, REEAgent.sense():

if (getattr(self.config, "harm_descending_mod_enabled", False)
        and self.beta_gate.is_elevated
        and new_latent.z_harm is not None):
    attn = getattr(self.config, "descending_attenuation_factor", 0.5)
    new_latent.z_harm = new_latent.z_harm * attn

Note: z_harm field holds z_harm_s (the sensory-discriminative stream). z_harm_a is in LatentState.z_harm_a and is not modified.

Validation Experiment (EXQ-325)

Design: Two conditions, 200 episodes each, 3 seeds.

CONTROL: harm_descending_mod_enabled=False (current behaviour)
DESCENDING: harm_descending_mod_enabled=True, descending_attenuation_factor=0.5

Both conditions use limb_damage_enabled=True (SD-022) to provide a genuine causal harm stream that is not fully predictable from position alone.

Metrics:

Metric	Target
`z_harm_s_norm_committed`	< `z_harm_s_norm_uncommitted` (within DESCENDING condition)
`z_harm_a_norm_committed`	Not significantly different from `z_harm_a_norm_uncommitted`
`benefit_ratio`	DESCENDING >= CONTROL (descending mod should not hurt navigation)
`n_committed_steps`	DESCENDING >= CONTROL (committed episodes should not shorten)

Pass criterion C1: z_harm_s is measurably lower during committed episodes than uncommitted. Pass criterion C2: z_harm_a shows no corresponding attenuation. Pass criterion C3: Task performance is maintained or improved (descending mod frees E3 from re-evaluating already-committed harm).

Note: Requires SD-020 (harm surprise PE) to have a well-shaped z_harm_s signal; the descending attenuation is most interpretable when z_harm_s encodes precision-weighted PE rather than raw accumulated harm.

Relationship to Other Claims

Claim	Relationship
SD-011	Provides the dual-stream split; z_harm_s and z_harm_a are separate targets
SD-020	z_harm_s should encode PE (not raw magnitude) before descending attenuation is meaningful
ARC-016	Precision weighting framework; attenuation factor could be precision-scaled in V4
MECH-090	BetaGate elevation identifies the committed state that triggers attenuation
MECH-220	Downstream claim: pain as control signal; SD-021 is the prerequisite substrate

Forward Notes

The current implementation applies a fixed scalar descending_attenuation_factor. A richer V4 implementation would modulate this factor dynamically via ARC-016 precision: alpha_eff = base_factor * (1 - precision_scale * current_precision / max_precision). Under high-precision (confident) commitment, attenuation would be stronger; under uncertain commitment, it would be lighter. This is the harm_nonredundancy_precision_scale analog for the descending pathway.