SD-051: safety_prediction.cue_specific_conditioned_inhibition_substrate

Claim ID: SD-051 Subject: safety_prediction.cue_specific_conditioned_inhibition_substrate Status: IMPLEMENTED Registered: 2026-05-04 Depends on: SD-011 (z_harm_a affective stream), SD-050 / MECH-302 (relief-completion teaching signal), MECH-094 (categorical tag write gate), MECH-057a (commitment-release pipeline) Blocks: MECH-304 (cue-specific conditioned safety prediction substrate)

Problem

REE agents learn to avoid harm but have no mechanism for learning that specific features of the world predict the absence of harm — the conditioned inhibition substrate. Without this, avoidance commitments cannot be released by safety cues; only harm cessation (MECH-302) or goal attainment (MECH-057a) close avoidance loops.

This is the architectural gap measured by the conditioned-inhibition discriminative-pair protocol (Andreatta 2012; Ng & Sangha 2023): a learned safety cue should suppress ongoing avoidance commitment even when a threat cue is concurrently present. Without SD-051, an REE agent cannot produce this behavioural signature.

Solution

A non-trainable ConditionedSafetyStore module that maintains a prototype representation of the world state at MECH-302 relief-completion event times (the EMA of z_world across all waking ticks when relief fires). At each step, it computes the cosine similarity of the current z_world to the stored prototype and returns a safety_prediction scalar.

Encoding (dorsal striatum / dlPFC analog): EMA prototype update on MECH-302 event ticks. z_safe_prototype = (1 - alpha) * z_safe_prototype + alpha * normalize(z_world) The prototype is the slowly-accumulated average of the z_world states that co-occurred with relief — the V3 analog of the dorsal striatal / dlPFC associative encoding identified in Laing et al. 2022 (conditioned inhibitor encoding, distinct from standard safety signal).

Per-step decay: z_safe_prototype *= (1 - decay_rate) (forgetting without reinforcement). Biologically motivated: conditioned inhibitors extinguish without continued pairing.

Expression pathway query (IL-to-CeA analog): cosine similarity between current z_world and z_safe_prototype is passed through a sigmoid with a gain parameter. safety_prediction = sigmoid(cosine_sim * gain) when ||z_safe_prototype|| > min_norm. Returns 0.0 when the store is empty (no events yet).

MECH-094 compliance: update() returns 0.0 and does NOT advance the prototype when sim_mode=True. Simulation and replay content cannot contribute to safety-cue learning or trigger safety-signal gates.

Data Flow 

CausalGridWorldV2.step() -> obs_dict
-> agent.sense() -> LatentStack.encode() -> latent.z_world [1, world_dim]
-> ConditionedSafetyStore.update(z_world[0], relief_event=_relief_completion_event, sim_mode=hypothesis_tag)
-> agent._conditioned_safety_signal: float (ephemeral, not in LatentState)

agent.select_action():
    # IL->CeA expression: lower avoidance commitment threshold when safety cue recognized
    if conditioned_safety_store and beta_gate.is_elevated:
        if _conditioned_safety_signal > safety_store_threshold:
            beta_gate.release(); _committed_step_idx = 0
            if valence_liking_enabled and not hypothesis_tag:
                residue_field.update_valence(z_world, VALENCE_LIKING,
                                             safety_store_commitment_weight, hypothesis_tag=False)
    _conditioned_safety_signal = 0.0  (cleared each tick)

Config Parameters

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

Parameter Default Purpose
use_conditioned_safety_store False master switch
safety_store_ema_alpha 0.1 prototype update rate per MECH-302 event
safety_store_decay_rate 0.001 per-step prototype decay (forgetting)
safety_store_min_norm 0.1 prototype L2 norm below which query returns 0
safety_store_threshold 0.5 cosine similarity threshold for commitment release
safety_store_commitment_weight 1.0 VALENCE_LIKING write magnitude on release

All parameters wired through REEConfig.from_dims().

Architecture Context

SD-051 is the encoding + expression substrate for MECH-304 (cue-specific conditioned inhibition). It parallels but does not duplicate MECH-303 (contextual passive safety representation, vmPFC/hippocampal substrate):

  • MECH-303: diffuse, context-level, no discrete teaching event. Accumulated non-harm in a context lowers background vigilance.
  • MECH-304 / SD-051: cue-specific, event-driven, MECH-302 teaching signal required. A discrete safety cue that co-occurred with relief becomes a targeted commitment-release trigger.

Laing et al. 2022 (7T fMRI, n=49) provides the load-bearing dissociation: conditioned inhibitors recruit dorsal striatum + dlPFC during encoding (SD-051 substrate); standard safety signals recruit vmPFC + hippocampus (MECH-303 substrate). The two pathways are anatomically distinct.

The IL-to-CeA expression pathway (Ng & Sangha 2023 Neuropsychopharmacol, Ng & Sangha 2023 Cereb Cortex) is the output gate: when the conditioned inhibitor is recognised, IL output suppresses CeA fear/avoidance output. In REE, this is modelled as safety_prediction → beta_gate release threshold.

What This SD Enables

  • MECH-304 substrate implementation (v3_pending cleared after V3-EXQ-519 validation)
  • Classical conditioned inhibition behavioural signature: learned safety cue suppresses avoidance commitment even in presence of concurrent threat cue
  • PTSD / chronic-avoidance failure mode prediction: if SD-051 is degraded but MECH-303 is intact, the agent feels safe in familiar environments but cannot use specific safety signals to calm down on demand – a substrate-level PTSD phenotype distinct from MECH-303 failure

V4-Deferred Extensions

The V3 substrate is minimal by design. Two extensions are deferred to V4:

  1. Approach attractor toward safety-signaling cues: V3 handles only commitment-release gating (output 1 of the MECH-304 spec). Output 2 — the safety cue becoming an approach attractor in its own right — requires V4 multi-step trajectory planning infrastructure (HippocampalModule completing V3 prerequisite gate, MECH-163). Without multi-step planning, a safety-cue approach attractor reduces to a step-wise proximity signal that cannot compete with harm avoidance commitments over trajectories. Registered in v3_v4_transition_boundary.md under V4 extensions.

  2. Contrastive cue-specific learning: The V3 EMA prototype conflates any z_world co-occurring with MECH-302 events, including contextual features that happen to be stable at relief time. A contrastive store (learning what distinguishes safety events from non-safety events via contrastive pairs) would produce sharper cue specificity. This requires a trainable encoder head and the phased-training infrastructure it entails. Substrate ceiling: V3’s non-parametric prototype may generalise the safety prediction too broadly in environments with stable non-safety features that happen to co-occur with relief. Registered in v3_v4_transition_boundary.md.

Module and Files

  • ree_core/safety/__init__.py – new package
  • ree_core/safety/conditioned_safety_store.py – ConditionedSafetyStore class (non-trainable; no nn.Module inheritance; pure arithmetic)
  • ree_core/utils/config.py – 6 config params in REEConfig, wired through from_dims()
  • ree_core/agent.py – import, __init__ instantiation, sense() tick, select_action() fire handler, reset() clear

Backward Compatibility

use_conditioned_safety_store=False by default. With this default:

  • agent.conditioned_safety_store is None
  • sense() skips the store update entirely
  • select_action() skips the safety-gate block entirely
  • No change to any existing experiment output

Biological Basis

Laing et al. 2022 (J Neurosci, 7T fMRI): conditioned inhibitors (safety cues learned via relief pairing) encode in dorsal striatum + dlPFC, dissociating from standard safety signals (vmPFC/hippocampus). Load-bearing dissociation supporting the two-system (MECH-303/304) architecture.

Ng & Sangha 2023 (Neuropsychopharmacology): IL->CeA pathway (chemogenetic dissection) is required for conditioned safety expression; IL->BLA is not required. Positions IL as the expression gate, not the encoding site.

Ng & Sangha 2023 (Cereb Cortex): IL single-unit recordings show cells that encode conditioned inhibitors specifically – the expression substrate has dedicated representational support.

Anatomical enrichment (2026-06-05, Silva et al. 2021)

Silva et al. 2021 (Nat Neurosci 24(7):964-974, DOI 10.1038/s41593-021-00856-y; PMID 34017129) – from the targeted_review_affect_stream_relief_safety_soothing lit-pull – adds two elements to MECH-304 (enrichment, not contradiction):

  1. A midline-thalamic transmission node: nucleus reuniens (NRe) -> basolateral amygdala (BLA). Combining viral tracing, fiber photometry, chemogenetics and closed-loop optogenetics in mice, Silva shows NRe registers and actively transmits safety signals to the BLA: NRe->BLA activity rises shortly before freezing cessation, pathway-specific inhibition impairs and activation facilitates the suppression of an established threat representation. This is the cleanest available causal, bidirectional demonstration that cue-specific safety prediction is its own functional system acting on the fear circuit (not low harm, not relief’s reward-offset reinforcement). In REE terms the reuniens relay is an additional node on the safety-expression pathway (parallel to IL->CeA) feeding the beta_gate commitment-release.

  2. Remote-vs-recent (time-since-encoding) dependence. The NRe->BLA relay is recruited specifically for the extinction of remote (systems-consolidated, ~30-day) aversive memories, not recent (~1-day) ones. This is a new functional axis orthogonal to the context-vs-cue axis separating MECH-303 from MECH-304: it indexes how consolidated the suppressed fear is. Recent-memory safety expression is handled by the existing MECH-303/304 machinery; suppressing an already-consolidated fear appears to require this distinct thalamo-amygdalar relay.

Candidate third safety sub-mechanism (flagged, not registered). The consolidation-age axis – “extinguishing an already-consolidated (remote) aversive memory requires a thalamo-amygdalar relay that recent-memory safety expression does not” – is a candidate for a future third safety sub-mechanism that would interface REE’s consolidation/replay machinery (MECH-092 SWR replay; sleep cluster) with the safety-expression pathway. It is deliberately not minted now: single rodent anchor (Silva 2021, conf 0.80), off the V3 critical path, no consolidation-interface substrate design. Promote only when (a) a second independent anchor lands and (b) a consolidation-state-conditional safety-expression design exists. Guards the over-split mirror of the philosophy-right/mechanism-wrong failure mode.

Neither the reuniens relay nor the remote/recent axis is implemented in the SD-051 ConditionedSafetyStore – this enrichment is documentation-only.

MECH-304 (cue-specific conditioned safety prediction – the claim this SD implements), MECH-303 (sister contextual pathway), MECH-302 / SD-050 (teaching signal source), SD-011 (z_harm_a affective stream), MECH-057a (commitment-release pipeline reused), MECH-094 (hypothesis-tag simulation gate)

Validation Experiment

V3-EXQ-519 (4-arm substrate-readiness diagnostic, queued 2026-05-04). See experiments/v3_exq_519_sd051_conditioned_safety_store_readiness.py.

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