MECH-303: Contextual Passive Safety Terrain

Claim ID: MECH-303 Subject: safety_prediction.contextual_passive_substrate Status: IMPLEMENTED 2026-05-04 Registered: 2026-05-03 Revised: 2026-05-04 (split from original MECH-303; cue-specific pathway moved to MECH-304 / SD-051) Depends on: SD-011, SD-012, ARC-007, MECH-304 (sister claim) Blocks: context-conditioned-vs-cue-conditioned safety extinction discriminative-pair experiment

Problem

The agent has no substrate for learning that an environment is generally safe. Without a contextual safety representation, avoidance commitments persist in familiar harmless environments, producing chronic vigilance that cannot be extinguished by passive exposure alone. This is distinct from MECH-304 (cue-specific conditioned safety), which requires explicit discrete relief events to update.

Biological grounding: Laing et al. 2022 (J Neurosci, 7T fMRI) dissociates two safety- learning systems – a standard safety signal pathway (contextual, passive accumulation, vmPFC + hippocampus) from a conditioned inhibitor pathway (cue-specific, event-driven, dorsal striatum + dlPFC). MECH-303 covers the vmPFC/hippocampal pathway. Kreutzmann 2020 (Psychopharmacology) shows IL is required for safety-signal expression. Meyer 2019 (PNAS) localises the contextual store to a vHipp-to-PL circuit, cross-species.

Solution

A parallel RBF field (safety_terrain_rbf_field) in ResidueField accumulates small per-step increments at the current z_world location whenever z_harm_a.norm() is below the quiescent threshold. Over repeated exposure to a harmless context, the safety terrain builds up a spatially-localised attractor at that context’s z_world signature.

In select_action(), the safety terrain is evaluated at the current z_world. When the scalar prediction exceeds contextual_safety_release_threshold AND the beta gate is elevated, the gate is released – lowering background avoidance commitment in familiar safe contexts.

Update rule (harm-absence accumulation)

sense() each step:
    harm_norm = z_harm_a.norm()   (SD-011 affective stream)
    if harm_norm < contextual_safety_harm_threshold   (harm is absent)
    AND hypothesis_tag == False                        (waking only, MECH-094)
        residue_field.accumulate_safety(z_world, contextual_safety_accum_weight)

The increment per step is small (default 0.01) so safety builds slowly, consistent with the slow/diffuse update rule of contextual safety learning (contrasted with the fast event-driven MECH-304).

Background vigilance gate

select_action() when beta_gate.is_elevated:
    safety_pred = residue_field.evaluate_safety(z_world)  [batch scalar]
    if float(safety_pred.mean()) >= contextual_safety_release_threshold:
        beta_gate.release()
        _committed_step_idx = 0
        _committed_anchor_keys = None

Architecture context

MECH-303 is the contextual/passive complement to MECH-304 (cue-specific conditioned inhibition, SD-051). The two systems dissociate behaviourally:

MECH-303: builds up over repeated safe context exposure (no discrete event needed); outputs a tonic reduction in background vigilance in that context.
MECH-304: updates on discrete MECH-302 relief-completion events; outputs a phasic commitment release when the stored safety cue is recognised.

The safety terrain follows the same RBF architecture as benefit_terrain (ARC-030) and the residue field itself, but uses a separate safety_terrain_rbf_field with its own accumulators – no sharing with VALENCE_LIKING or benefit_terrain.

Implementation surface

New module file: None – extension to ree_core/residue/field.py.

ResidueField changes:

safety_terrain_enabled attribute from ResidueConfig.safety_terrain_enabled
safety_terrain_rbf_field: RBFLayer instantiated when enabled
total_safety: Tensor buffer (total accumulated safety)
num_safety_steps: Tensor buffer (count of harm-absent steps accumulated)
accumulate_safety(z_world, magnitude, hypothesis_tag) – write path
evaluate_safety(z_world) – read path, returns [batch] scalar

All defaults are no-op. from_dims(use_contextual_safety_terrain=True) auto-enables residue.safety_terrain_enabled.

What this enables

V3-EXQ-520: 4-arm substrate-readiness diagnostic
- ARM_0 OFF: bit-identical baseline
- ARM_1 write-only (release_threshold=999): terrain accumulates, no gate fires
- ARM_2 full MECH-303 (release_threshold=1.0): gate fires in harm-free contexts
- ARM_3 gate-only (accum_weight=0): terrain empty, no gate – equals ARM_0
Future: context-conditioned-vs-cue-conditioned safety extinction discriminative-pair experiment (MECH-303 ablated: extinction slower for contextual safety; MECH-304 ablated: extinction slower for cue-specific safety)

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

The targeted_review_affect_stream_relief_safety_soothing lit-pull (2026-06-05) surfaced Silva et al. 2021 (Nat Neurosci 24(7):964-974, DOI 10.1038/s41593-021-00856-y), which causally establishes a midline-thalamic relay – nucleus reuniens (NRe) -> basolateral amygdala (BLA) – that registers and actively transmits safety signals onto the fear circuit (NRe->BLA activity rises before freezing cessation; pathway-specific inhibition impairs and activation facilitates extinction).

This is enrichment, not contradiction. Two points apply to MECH-303:

Cluster-level anatomy. The safety cluster’s expression machinery is prefrontal-hippocampal-thalamic, not prefrontal-hippocampal alone. Silva’s causal evidence bears most directly on the sister claim MECH-304 (the active, cue-specific safety signal that acts on the fear circuit), so MECH-303 carries only the lighter cluster-level note that the contextual safety representation’s downstream expression shares this thalamo-amygdalar relay. The MECH-303 vHipp-to-PL contextual store + IL expression anatomy stands unchanged.
Remote-vs-recent (time-since-encoding) dependence. Silva’s NRe->BLA relay is recruited for the extinction of remote (systems-consolidated, ~30-day) aversive memories and not recent (~1-day) ones. This indexes how consolidated the suppressed fear is – a new axis orthogonal to the context-vs-cue axis that separates MECH-303 from MECH-304. It is tracked on MECH-304 as a candidate third safety sub-mechanism (consolidation-age-conditional safety expression), deliberately not minted as a claim yet (single rodent anchor, off the V3 critical path, no consolidation-interface design). See MECH-304 notes.

Neither the thalamic relay nor the remote/recent axis is implemented in the SD-052 substrate – this enrichment is documentation-only.

MECH-303 (this claim): contextual_passive_substrate
MECH-304 / SD-051: cue_specific_conditioned_inhibition_substrate (sister, already implemented)
MECH-302 / SD-050: relief_completion_event (teaching signal for MECH-304; harm source for MECH-303)
SD-011: dual nociceptive streams (z_harm_a is the harm stream MECH-303 monitors)
SD-012: homeostatic drive (anchors “safety” as drive-baseline-not-rising)
ARC-007: hippocampal path memory (contextual encoding substrate)
ARC-030 / MECH-117: benefit terrain (parallel architecture; separate RBF field)
MECH-094: hypothesis tag (blocks safety accumulation during simulation/replay)