SD-050: relief.suffering_derivative_comparator

Claim ID: SD-050 Subject: relief.suffering_derivative_comparator Status: IMPLEMENTED Registered: 2026-05-04 Depends on: SD-011 (z_harm_a affective stream), MECH-057a (commitment-release pipeline), MECH-091 (salient-event phase-reset), MECH-094 (categorical tag write) Blocks: MECH-302 (relief-completion event substrate), MECH-303 (safety-cue prediction teaching signal)

Problem

REE’s goal-completion pipeline (MECH-057a beta-gate drop + MECH-091 phase-reset + MECH-094 VALENCE_LIKING write) fires on goal-attainment events. There is no architecturally analogous event for the end of a period of suffering: avoidance commitments accumulate during harm, but no clean “relief completion” event closes the episode and tags what just worked. This leaves MECH-302 untestable – the claim that the same downstream pipeline reuses for sustained suffering-reduction cannot be verified without a comparator that detects the crossing.

Solution

A non-trainable rolling-window descent detector (SufferingDerivativeComparator) on z_harm_a.norm(). On each waking tick (gated by simulation_mode=False), the comparator appends the current norm to a fixed-length FIFO buffer. When:

1. The buffer is full (window_length ticks of data)
2. The initial (oldest) norm exceeds min_initial_norm (prevents false fires on a stream that was already quiet)
3. The drop from initial to final exceeds drop_threshold

the comparator returns True and agent.sense() sets _relief_completion_event = True. This flag is consumed and cleared by select_action() in the same tick (adjacent to the MECH-091 urgency block, opposite polarity).

On event fire (select_action()):

• If beta is elevated: beta_gate.release() + _committed_step_idx = 0 (MECH-057a reuse)
• If valence_liking_enabled and not hypothesis_tag: ResidueField.update_valence(z_world, VALENCE_LIKING, relief_completion_weight) (MECH-094 reuse)

Architecture Context

SD-050 is architecturally adjacent to MECH-091 (urgency interrupt). MECH-091 fires on an UPWARD z_harm_a spike – the “pain is suddenly worse, abort” signal. SD-050 fires on a DOWNWARD z_harm_a descent – the “sustained relief, episode over” signal. The two share the downstream pipeline by design: both are salient events whose polarity is set at the input (nociceptive escalation vs nociceptive de-escalation) but whose consolidation machinery is shared (commitment release + valence write).

The comparator reads z_harm_a (C-fiber / affective-motivational stream, SD-011), NOT z_harm (A-delta / sensory-discriminative stream). This is biologically grounded: Navratilova 2012 shows pain relief activates VTA-DA + NAc-shell DA via the same circuit as reward; the relief signal is carried by the affective stream, not the discriminative one.

The MECH-094 gate (hypothesis_tag) prevents simulation and replay content from triggering relief-completion writes: only waking-path z_harm_a observations advance the comparator.

Data Flow

CausalGridWorldV2.step() -> obs_harm_a [harm_obs_a_dim]
-> agent.sense(obs_harm_a=...) -> AffectiveHarmEncoder -> z_harm_a [1, z_harm_a_dim]
-> z_harm_a.norm() -> SufferingDerivativeComparator.tick(norm, simulation_mode=False)
-> agent._relief_completion_event (bool, set in sense(), consumed in select_action())
-> select_action():
     if relief_completion_event and beta_gate.is_elevated:
         beta_gate.release(); _committed_step_idx = 0
     if relief_completion_event and valence_liking_enabled and not hypothesis_tag:
         residue_field.update_valence(z_world, VALENCE_LIKING, relief_completion_weight)
     _relief_completion_event = False  (cleared each tick)

Config Parameters

All gated by use_suffering_derivative_comparator=False (default, no-op):

Parameter	Default	Purpose
use_suffering_derivative_comparator	False	master switch
suffering_window_length	5	rolling-window FIFO depth (ticks)
suffering_drop_threshold	0.10	required drop from window[0] to window[-1]
suffering_min_initial_norm	0.05	floor on window[0] before detection is active
relief_completion_weight	1.0	VALENCE_LIKING write magnitude on event fire

All parameters wired through REEConfig.from_dims().

What This SD Enables

• MECH-302: relief-completion event substrate (the comparator IS the MECH-302 substrate)
• MECH-303 teaching signal: the MECH-302 event is the signal that updates MECH-303’s safety-cue predictive store
• Chronic-anxiety / persistent-avoidance failure mode: with comparator OFF, avoidance commitments accumulate without tag-and-release (no VALENCE_LIKING write on avoidance completion); PTSD/anxiety analog substrate

Module and Files

• ree_core/comparator/suffering_derivative_comparator.py – SufferingDerivativeComparator class
• ree_core/comparator/__init__.py – package (shared with TPJComparator)
• ree_core/utils/config.py – 5 config params in REEConfig, wired through from_dims()
• ree_core/agent.py – import (line ~93), __init__ instantiation (line ~328), sense() tick (line ~1801), select_action() fire handler (line ~2094), reset() clear (line ~1061)

Backward Compatibility

use_suffering_derivative_comparator=False by default. With this default:

• agent.suffering_comparator is None
• sense() skips the comparator tick entirely
• select_action() skips the relief-completion block entirely
• No change to any existing experiment output

All 5 smoke tests PASS 2026-05-04 (backward compat, feature ON 10 steps, comparator logic, flat-signal no-fire, min_initial_norm guard).

Biological Basis

Navratilova et al. 2012 (PNAS): pain relief activates VTA dopamine and NAc-shell dopamine via the same circuit as appetitive reward; blocked by DA antagonists and produces conditioned place preference – confirming shared machinery. Tanimoto 2004 (Nature): Drosophila opponent-timing confirms relief as a positive reinforcer when paired with appropriate temporal structure. Andreatta 2012: relief from fear -> striatum (not amygdala), mirroring the goal-attainment vs harm-anticipation dissociation. The suffering-derivative comparator (not a pooled DA readout) is the architecturally correct trigger per Bromberg-Martin 2010 DA-heterogeneity qualifier: value-coding VTA-DA is the substrate, but the REE trigger must be the derived comparator to avoid mixing salience-coding DA.

Related Claims

MECH-302 (relief-completion event), MECH-303 (safety-cue prediction teaching signal), SD-011 (z_harm_a affective stream source), MECH-057a (commitment-release reused), MECH-091 (opposite-polarity urgency interrupt), MECH-094 (VALENCE_LIKING write gate)

Validation Experiment

V3-EXQ-515 PASS 2026-05-04 (4-arm comparator logic diagnostic: unit tests on synthetic norm sequences). V3-EXQ-516 queued (4-arm agent-loop integration diagnostic: OFF / event-fires / valence-write / flat-signal). See experiment script experiments/v3_exq_516_mech302_suffering_derivative_integration.py.