Trainable relief / safety affordance learners (MECH-375 / MECH-376 / Q-067)

Status: candidate cluster, registered 2026-06-09. Home doc for the trainable successors to the fixed escape-affordance bridge.

Scope: V3-NARROW minimal trainable bridge (a post-V3-EXQ-603i successor candidate, close to the V3 critical path). A V4 / V5-rich tier is described at the end and is kept off the V3 path.

Source thought: docs/thoughts/2026-06-08_Trainable_Releif_and_Safety.md

Why this cluster exists

The escape-affordance bridge (SD-059 / MECH-358) currently binds relief and safety to escape via a fixed-arithmetic, per-first-action-class EMA credit table. The relief-completion machinery it consumes (MECH-302 / SD-050) is a non-trainable rolling-window comparator that detects a sustained z_harm_a descent on the current state. The safety substrate (MECH-303 contextual passive store, MECH-304 cue-specific conditioned inhibition) is a pair of substrate hypotheses whose implemented V3 form is, again, the fixed EMA table.

None of these learns a generalising function of state x action x (threat / cue) context. V3-EXQ-603h showed the consequence: the agent could suppress freezing and act under threat, but did not learn that a specific action / location / policy was the way out – a scalar avoidance efficacy did not bind relief to an escape direction. The biologically stronger architecture makes both relief and safety trainable.

What REE already owns (cross-ref, NOT duplicated)

Owned Role Why the trainable head is not a duplicate
SD-050 / MECH-302 non-trainable suffering-derivative comparator + relief-completion pipeline reuse DETECTS relief on the current state; does not bind / generalise credit to an action
MECH-303 / MECH-304 contextual + cue-specific safety substrate hypotheses substrate hypotheses; implemented form is a fixed EMA table, not a trained predictor
SD-058 / MECH-357 instrumental-avoidance acquisition + ilPFC freeze-suppression (scalar efficacy) scalar global efficacy; penalises freeze, does not pick a direction
SD-059 / MECH-358 escape-affordance bridge (fixed per-action-class EMA credit table) fixed arithmetic; no generalising function over continuous state / action / context

The two trainable heads

MECH-375 – Trainable relief critic Q_relief(state, action, threat_context)

A learned parametric head predicting expected harm / suffering reduction after an action under threat. Function = negative-reinforcement credit assignment (“this action reduced suffering / terminated harm”), distinct from generic reward / wanting even if it reuses reward machinery.

  • Inputs (candidate): z_world, z_self, threat / harm streams (esp. z_harm_a), action class / candidate first action, possibly local E3 trajectory features.
  • Training target: positive when a directed action under threat is followed by a z_harm_a drop; stronger when temporally close, action-contingent, and not explained by passive drift; negative / extinction when the expected relief action fails.
  • Role: under future threat, bias E3 toward actions predicted to reduce harm; silent when safe; never credits no-op / freeze unless explicitly modelling passive safety.

MECH-376 – Trainable safety predictor P_safety(state, cue, action, context)

A learned parametric prospective predictor of threat-absence / response-produced safety / conditioned inhibition. Licenses commitment-release, approach, and recovery. Distinct from z_goal (predicts threat absence, not goal presence) and from relief (prospective prediction, not an aversive-offset event).

  • Inputs (candidate): z_world, cue / context features, action class / recent action, threat history, time since threat offset, possibly a hippocampal context slot / rule representation.
  • Training target: positive when a cue / context / action predicts threat absence; negative when a “safe” cue is followed by threat recurrence; contrastive so safety cannot collapse to “low harm” or “recent relief”.
  • Role: support commitment release, permit approach to safety affordances, stabilise recovery, prevent persistent defensive mode when threat is genuinely absent.

Why both (Q-067)

Relief alone teaches “that action reduced harm” but need not produce a stable prospective safety model; safety alone teaches “this context / cue is safe” but need not solve action-specific escape under active threat. The mature avoidance system needs action-contingent relief learning (MECH-375) for escape and cue / context-contingent safety learning (MECH-376) for recovery / inhibition / future approach. Q-067 is the decomposition adjudicator routed by the V3-EXQ-603i outcome (partial pass -> train the missing half first; fail-with-non-vacuity -> fixed table too crude, trainable learner required; credit-never-fires -> route back to substrate readiness).

Design guardrails (binding on any implementation)

The trainable system must remain:

  • bounded – cannot dominate E3 scoring (bias_scale clamp, as on MECH-358);
  • threat-gated – no global approach bias when safe;
  • extinguishable – failed relief / safety predictions decay;
  • contrastive – safety cannot mean “low harm” or “recent relief”;
  • action-bound – relief credits the action / policy that produced harm reduction;
  • cue / context-bound – safety credits the cue / context / action that predicts threat absence;
  • MECH-094-safe – simulated relief / safety in hypothesis mode must not train the waking learner unless explicitly authorised as play / training.

Governance scoping

All three claims are status: candidate, epistemic_category: substrate_conditional, implementation_phase: v3 (v3_pending: true). The substrate_conditional category is deliberate: the trainable parametric heads are planned but not yet built (the current bridge is the fixed table), so promote / demote is suppressed and the cluster is kept off the IGW experiment-proposal lane until a head exists. The right next step is to decide whether to build the minimal trainable bridge (a later step), not to queue a vacuous probe on the fixed-table substrate.

V4 / V5-rich tier (OFF the V3 path – do not build in V3)

The thought’s richer architecture is explicitly deferred and not registered as part of the V3-narrow cluster:

  • expression-as-action-geometry for relief / safety;
  • richer continuous state / policy / location indexing (beyond first-action class);
  • hippocampal affect-gradient indexing of relief; social safety attribution;
  • multi-cue / rule-representation safety contexts.

These belong to the V4 / V5 enrichment programme and must not enter the V3 critical path; they are recorded here only so the boundary is explicit.

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