Rule Apprehension Layer (architecture anchor)

Status: registered architectural slot. Weak reading (ARC-062) at implementation_phase=v3 candidate, BLOCKED (GAP-B status=blocked; 543f/543g/543h falsifier cluster non_contributory, epistemic_category=substrate_ceiling; weak-route verdict in flight via V3-EXQ-543i /failure-autopsy + V3-EXQ-543j cross-machine confirm). Strong reading (ARC-063) brought forward 2026-05-18 from implementation_phase=v4 to v3 candidate (user-directed, independent-need rationale). New socially-scaffolded rule-population pathway ARC-077 (+ children MECH-337 / MECH-338) registered 2026-05-18 as structural slots, hard-gated on the GAP-L biology lit-pull. Falsifiable claim: MECH-309 (monomodal collapse as equilibrium without rule-apprehender).

ML translation (for readers from an ML background). “Rule apprehension” is the policy layer’s capacity to form and hold multiple distinct behavioural regimes rather than collapse to a single averaged mode under gradient descent – the unimodal-policy mode-collapse / mode-averaging failure (a Gaussian policy facing a go-left-or-right choice picks the catastrophic average; the canonical motivation for multimodal policies – MDNs, diffusion policies, IBC). The V3 weak-reading instantiation (ARC-062, gated_policy) is structurally a context-gated two-expert mixture-of-experts: two scoring heads share encoder features and a learned context discriminator emits a soft gate w in [0,1]; the nearest RL framing for the regimes themselves is the options framework (Sutton 1999; see the targeted_review_rule_apprehension_vocabulary_mapping lit-pull). The distinctive claim is not “use an MoE.” MECH-309 asserts the regimes must be created by a non-Bayesian rule-creator, not merely weighted by a trainer that refines pre-given ones – which is exactly why a vanilla end-to-end MoE, itself prone to router/expert collapse (the same equilibrium MECH-309 diagnoses), is necessary-but-not-sufficient. The cognitive-science word “apprehension” is retained deliberately to mark that creation-vs-selection distinction; the term is not renamed to a standard ML label precisely because the standard labels flatten it.

Origin: this document was promoted from docs/thoughts/2026-05-04_Waking_rule_apprehension_later_sleep_schema.md to docs/architecture/rule_apprehension_layer.md on 2026-05-08 after the SD-054 substrate-purpose-validation discussion identified the rule-apprehension layer as the proximate next-stage architectural commitment that explains the EXQ-433e/f / 523-series non_contributory pattern. The body of the doc is preserved verbatim from the original intake; this preamble adds the cross-reference scaffolding.

Cross-references:

MECH-309 (claims.yaml) – the diagnostic claim (“monomodal collapse is the equilibrium of a parametric-policy agent without a rule-apprehension layer”). MECH-309 is falsifiable by the ARC-062 weak-reading experiment.
ARC-062 (claims.yaml) – weak-reading architectural slot: gated-policy architecture with learned context discriminator. V3 first pass; engineering-cheap.
ARC-063 (claims.yaml) – strong-reading architectural slot: distributed CandidateRule field with tolerance-gated availability, hippocampal-rollout-eligibility shaping, structured per-action evidence-trace records, waking-vs-sleep refinement asymmetry. Brought forward 2026-05-18 (implementation_phase v4 -> v3, v3_pending false -> true). Now the field that the ARC-077 population pathway fills.
ARC-077 (claims.yaml) – socially-scaffolded rule-population pathway (registered 2026-05-18). The third rule-source, peer to ARC-062 (top-down) and ARC-064 (bottom-up), upstream-fed by ARC-065 (diversity). Caregivers populate the candidate-rule space from outside the agent. Structural slot; hard-gated on the GAP-L biology lit-pull + a caregiver-agent substrate that does not exist in V3.
MECH-337 (claims.yaml) – caregiver-scaffolded rule-population mechanism (child of ARC-077; the “in” face of the ARC-063 field).
MECH-338 (claims.yaml) – cue-driven context-bound rule retrieval (child of ARC-077; the “cue pipeline” / “select” face of the ARC-063 field).
ARC-064 / ARC-065 (claims.yaml) – peer bottom-up rule-discovery route / upstream behavioural-diversity-generation pathway. ARC-077 is the third member of the rule-source triad they belong to.
SD-054 (claims.yaml + sd_054_reef_enrichment_substrate.md) – the substrate that exposes the gap between substrate-readiness and substrate-purpose-validation under trained policy.
SD-029, MECH-256, MECH-269 – the downstream measurement and representational precondition cluster.
SD-033a (sd_033a_lateral_pfc_analog.md) – the committed rule_state substrate; sits downstream of the apprehension layer on the apprehension -> commitment pipeline.
SD-034 (sd_034_governance_closure_operator.md) – closure operator on the release side; complementary to ARC-063.

Socially-scaffolded rule-population pathway (ARC-077) – 2026-05-18 bring-forward

Trigger. The ARC-062 weak route is blocked: GAP-B status=blocked; the 543f/543g/543h falsifier cluster is non_contributory with epistemic_category=substrate_ceiling; ARC-062 / MECH-309 carry narrow_supports_flag (zero reliable contributory trained-policy evidence). The weak-route verdict is still in flight (V3-EXQ-543i /failure-autopsy of the diff-OFF non-reproduction + the concurrent V3-EXQ-543j cross-machine confirmation, ahead of ARC-062/MECH-309 governance). On the user’s direction and on the independent-need rationale (ARC-063 is required regardless of the weak-route verdict for moral-residue attribution and clinically-realistic failure modes), the strong reading was brought forward and a third rule-source pathway registered. ARC-062 posture is unchanged by this work.

Why a new peer pathway, not a sub-mechanism of ARC-063 (framing decision, user delegated “you decide from the architecture”). ARC-063 is the field – the representation that HOLDS candidate rules. MECH-309’s whole point is that gradient/Bayesian learners do not invent the hypothesis space, so “where do candidate rules COME from” is the architecturally primary question, and the cluster already partitions rule-sources by pathway: ARC-062 top-down (gradient pressure finds a cut), ARC-064 bottom-up (compress regularities out of own behaviour), ARC-065 upstream (behavioural diversity to cluster over). The biologically dominant source is none of these: caregivers populate the infant’s candidate-rule space from outside via scaffolded exposure. That is a third source, peer to ARC-062 / ARC-064, that feeds the ARC-063 field – so it gets its own ARC-level slot (ARC-077), not a slot inside ARC-063.

Three faces of the ARC-063 field, made explicit:

population (“in”) – MECH-337: a caregiver/teacher agent structures exposure within the zone of proximal development and ostensively marks a regularity; it is written into the ARC-063 field as a CandidateRule the agent could not have invented alone, tagged with the scaffolding context. Population is context-bound by construction – this is what “context-bound rule population” means.
select (“the cue pipeline”) – MECH-338: environmental/internal cues are matched against CandidateRule context tags to gate rollout-eligibility (context-dependent / cued recall; encoding-specificity). A rule scaffolded in context C is dormant until C, or a learned cue for C, recurs. Feeds – does not replace – the ARC-063 Tolerance-Principle availability gate.
express (“out”): hippocampal-rollout-eligibility shaping. This is already an ARC-063 commitment (rules bias which futures are easy to imagine); ARC-077 cross-references it rather than re-registering it.

HARD PREREQUISITE – GAP-L biology lit-pull (NOT yet discharged). Per the standing biology-before-formal-definitions rule, ARC-077 / MECH-337 / MECH-338 are registered as structural slots only (lit_conf=0, must not be promoted past candidate). No implementation (V3 or V4) may begin until a caregiver-scaffolding / cued-recall / context-dependent rule-acquisition lit-pull is discharged: Csibra & Gergely natural pedagogy / ostensive cueing; Vygotsky ZPD + Wood/Bruner/Ross scaffolding; Tomasello shared intentionality; joint attention / social referencing; Godden & Baddeley context-dependent memory; Tulving & Thomson encoding-specificity. A second hard gate: the pathway needs a caregiver/teacher-agent substrate, which does not exist in V3 (single-agent). See arc_062_rule_apprehension_plan.md GAP-L and the 2026-05-18 decision-log entry. implementation_phase=v3 here means “active V3 design target under the V3-Pending Gate”, not “V3-implementable now”.

Strong-reading-internal sub-claim landscape (corrected 2026-05-18)

The original 2026-05-08 placeholder list assigned MECH-310..313 and is now partly stale – two of those IDs were reused for unrelated claims and MUST NOT be reused for the strong reading:

tolerance-gated rule-availability mechanism – placeholder was “MECH-310”; MECH-310 still free, allocate at V4 take-up.
rules-as-rollout-selection-structures mechanism – placeholder was “MECH-311”; MECH-311 still free.
rule-evidence-trace records mechanism – placeholder “MECH-312” is NOW TAKEN (policy.rule_arbitration_multi_variable); allocate a fresh ID.
waking-vs-sleep rule-refinement asymmetry mechanism – placeholder “MECH-313” is NOW TAKEN (policy.stochastic_noise_floor_lc_ne_tonic_analog); allocate a fresh ID.
INV: action-requires-collapse (INV-COMMITMENT-001 candidate in intake).
INV: no-waking-rule-is-final (INV-UNCERTAINTY-001 candidate).

These strong-reading-internal mechanisms stay unregistered to avoid premature commitment. The ACTIVE registered sub-cluster is the socially-scaffolded population triad above (ARC-077 / MECH-337 / MECH-338). The intake-doc-original “Proposed REE claims” section below preserves the original intake labels (MECH-TOLERANCE-ACCESS-001 etc.); those are intake-doc internal design tokens and do not match REE’s MECH-NNN registry – treat as design tokens, not registered IDs.

Thought Intake: Tolerance-Gated Rule Apprehension, Hippocampal Rollout Selection, and Basal Ganglia Commitment

Intake title

Rules shape possible futures; goals weight futures; basal ganglia commits futures into action

⸻

Source / trigger

User-linked paper:

“A simple threshold captures the social learning of conventions” Proceedings of the National Academy of Sciences (PNAS) Digital Object Identifier: 10.1073/pnas.2508061123

The paper appears relevant because it proposes that convention learning may involve a threshold-like transition from probabilistic sampling to stable rule use. The key REE value is not only social convention learning, but the possibility of a general mechanism by which observed regularities become usable rules despite exceptions.

⸻

User-originating insight

Direct user-originating fragments from this exchange:

“This could well be part of basal ganglia gating for quick and easy rule apprehension during waking.”

“Sleep based refinement of the buckets would then have a starting point with clear evidence record of the ‘experiments’ of the rules apprehended during waking.”

“Of course the tolerance principle could easily also be instantiated by interactions between goal systems and hippocampus and there may be less left to sleep than the above implies.”

“The basal ganglia commitment which must by its very nature collapse probabilities into a decision is where the division crystallises but which rules were being used for the decision perhaps seems more like other systems.”

“So the rules must be contained within the hippocampal role out and the approach goal or avoid thing. Or at least between the interaction between what leads to the roll out choice.”

Authorship note: “role out” above is preserved as typed, but the intended term appears to be rollout.

⸻

Central correction to earlier framing

The initial basal-ganglia-and-sleep framing was useful but too simple.

A better REE formulation is:

The Tolerance Principle may describe a functional threshold by which candidate regularities become usable rules. This threshold should not be prematurely localised to basal ganglia gating alone.

Instead:

Rules are likely constructed, represented, and made available through interactions between hippocampal rollout generation, goal systems, cortical predictive models, affective/salience systems, and memory structures. Basal ganglia gating is where the selected rule-weighted rollout crystallises into action.

Sleep/offline processing remains important, but it should not be framed as the only place where rule refinement happens. Waking cognition may already perform substantial rule testing, weakening, splitting, and re-bucketing.

⸻

Core REE claim

The key architecture is:

regularity detection → tolerance-gated rule availability → hippocampal rollout biasing → approach/avoid goal weighting → basal ganglia commitment → action → evidence record → waking and sleep/offline refinement

This gives REE a candidate route from probabilistic observation to embodied decision without requiring premature certainty.

⸻

Conceptual model

Regularity detection

The agent encounters repeated patterns across experience.

Examples:

This cue often predicts reward. This social tone often predicts rejection. This context often predicts danger. This action often produces relief. This person often responds well to repair. This rule has many exceptions.

At this stage, these are not yet committed beliefs. They are candidate regularities.

⸻

Tolerance-gated rule availability

A regularity becomes rule-like when the system has enough supporting evidence relative to exceptions.

The Tolerance Principle can be treated as a candidate mathematical form for this:

A rule becomes productive when exceptions remain tolerable relative to the size of the observed set.

REE should not assume the exact equation is universally implemented in the brain. The important architectural idea is:

Some exceptions can be tolerated without preventing rule formation, but too many exceptions should prevent the regularity from becoming a productive rule.

So the Tolerance Principle may function as a rule availability gate, not necessarily a final decision gate.

⸻

Hippocampal rollout selection

The rule then affects which futures are generated or prioritised.

The hippocampal system is not simply retrieving memories. It is generating possible trajectories:

What happens if I approach? What happens if I avoid? What happens if I wait? What happens if I ask? What happens if I repair? What happens if I challenge? What happens if I submit?

Rules may be embedded in the rollout landscape itself.

A rule may become psychologically active when it begins shaping which futures are easy to imagine.

Example:

similar_person + similar_tone + prior_harm → threat rollout becomes highly available

Another:

familiar_safe_context + affiliative_signal → approach rollout becomes highly available

Another:

many_exceptions_to_threat_rule → ambiguity / information-seeking rollout becomes available

So the rule is not merely:

X predicts Y

It is closer to:

In this context, X makes this kind of future worth simulating.

⸻

Goal systems and approach/avoid weighting

Goal systems then weight candidate rollouts according to motivational and bodily relevance.

Relevant weighting dimensions may include:

reward threat attachment status curiosity information gain energy cost social cost moral cost bodily state urgency pain fatigue shame hope

This gives each rollout an action tendency:

approach avoid freeze submit explore repair delay escalate inhibit ask withdraw

The rule may therefore be located not in one system, but in the interaction between:

which futures are generated + which futures are motivationally weighted + which futures are suppressed or amplified

⸻

Basal ganglia commitment

The basal ganglia is then where the weighted field must pass through a bottleneck.

The rule is not necessarily born there.

Instead:

The basal ganglia crystallises a distributed rule-weighted probability field into action, inhibition, sequence, or delay.

The organism may internally preserve uncertainty, but the body cannot act in superposition. An action, non-action, inhibition, or delay must be selected.

This is where:

probability field → committed behavioural trajectory

The decision crystallises here because action requires collapse.

⸻

Evidence record

Once a rule-weighted rollout is committed into action, the system should preserve an evidence record.

This evidence record should include:

which rule-like regularities were active which rollouts were generated which rollouts were suppressed which goal weights were applied which action was selected what outcome occurred prediction error reward harm relief social feedback moral residue exception evidence contextual modifiers

This is crucial because later refinement requires knowing not only what happened, but what rule was being tested by the action.

Waking rule-use becomes an experiment.

⸻

Waking refinement

Waking cognition may already perform substantial refinement.

It may:

weaken a rule strengthen a rule split a rule merge rules notice exceptions detect hidden subgroups update confidence change rollout eligibility alter goal weights suppress overgeneralisation promote useful local rules retire failed rules

This correction matters because sleep should not be treated as the only site of bucket refinement.

⸻

Sleep/offline refinement

Sleep/offline processing can still play a major role, but as continuation, compression, renormalisation, and wider reorganisation.

Sleep may:

compress evidence renormalise affective weights replay rule-action-outcome sequences identify exception clusters split overbroad buckets merge redundant schemas prune failed rules stabilise useful schemas reduce noise recontextualise trauma-weighted rules support counterfactual recombination alter future rollout eligibility

Thus sleep does not begin refinement from nothing. It receives waking-generated evidence records and partially refined rule structures.

Better formulation:

Sleep/offline processing continues and reorganises refinement rather than being the sole adjudicator of waking rule experiments.

⸻

REE anatomical / functional mapping

REE function Likely system contribution Candidate regularity detection cortical predictive systems, hippocampal comparison, memory systems Exception tracking hippocampus, prediction-error systems, cortical conflict monitoring Rule availability threshold distributed tolerance-like function Rollout generation hippocampal system Rollout selection bias hippocampus interacting with goals and salience Approach/avoid weighting goal systems, affective systems, interoception, salience networks Action commitment basal ganglia gating Action sequence selection basal ganglia, motor and associative loops Evidence record hippocampal-cortical tagged memory Waking update hippocampal-cortical-goal interaction Sleep refinement offline replay, consolidation, affective renormalisation, schema update

⸻

Key distinction

Earlier phrasing:

Basal ganglia apprehends rule. Sleep refines rule.

Revised phrasing:

Distributed systems construct and make rules available. Hippocampal rollout systems embed rules into possible futures. Goal systems weight those futures. Basal ganglia crystallises the weighted field into action. Waking and sleep systems both refine the rule landscape.

⸻

Important REE sentence candidates

Rules shape possible futures; goals weight futures; basal ganglia commits futures into action.

The basal ganglia is not necessarily where the rule is born; it is where competing rule-weighted futures must pass through the narrow gate of action.

The rule-field is distributed; the commitment bottleneck crystallises in basal ganglia gating.

A rule may become psychologically real not when it is verbally endorsed, but when it starts constraining which futures the system can easily imagine.

Waking rule-use is an experiment, and the evidence record of that experiment must be preserved for later refinement.

Sleep does not start from raw experience alone. It receives the experimental record of rules waking cognition tried to use.

The transition from uncertainty to rule-use is not the same as the transition from uncertainty to truth.

⸻

Proposed REE claims

MECH-TOLERANCE-ACCESS-001 — Tolerance threshold as rule-access mechanism

Tolerance-like thresholds may determine when a candidate regularity becomes available as a productive rule for decision systems. The threshold does not need to be the final action-selection mechanism.

Status: candidate mechanism.

⸻

MECH-HIP-GOAL-RULE-001 — Rules as rollout-selection structures

In REE, rules should be represented not merely as explicit propositions but as structures that bias which hippocampal rollouts are generated, selected, amplified, or suppressed under goal-weighted approach/avoid conditions.

Status: candidate mechanism.

⸻

ARC-HIP-GOAL-BG-001 — Rollout-weighting-commitment pathway

Hippocampal systems generate candidate futures. Goal and salience systems weight those futures according to approach, avoidance, threat, reward, attachment, energy, social meaning, and moral cost. Basal ganglia gating crystallises the weighted field into committed action, inhibition, sequence, or delay.

Status: architectural commitment candidate.

⸻

ARC-BG-COMMIT-001 — Basal ganglia as action-collapse gate

The basal ganglia should be modelled as a commitment gate that helps collapse probabilistic, competing candidate policies into action, inhibition, or selected sequence. It need not be the primary site of rule discovery.

Status: architectural hypothesis.

⸻

MECH-RULE-FIELD-001 — Distributed candidate rule field

REE should represent pre-action rule-use as a distributed field of candidate rules, heuristics, predictions, goals, and affectively weighted priors arising from hippocampal, cortical, goal, social, interoceptive, and salience systems.

Status: candidate mechanism.

⸻

MECH-EVIDENCE-TRACE-001 — Rule evidence records

Every provisional rule used in action selection should preserve a structured evidence record, including exemplars, exceptions, generated rollouts, selected action, suppressed alternatives, affective weighting, prediction error, action outcomes, and downstream harm/benefit signals.

Status: implementation-relevant mechanism.

⸻

ARC-WAKE-REFINE-001 — Waking refinement is substantial

Waking cognition should be modelled as already capable of substantial rule refinement. Goal-directed cognition and hippocampal comparison may update, split, weaken, or recontextualise candidate rules before sleep.

Status: architectural correction.

⸻

ARC-SLEEP-REFINE-001 — Sleep continues rather than originates refinement

Sleep/offline processing should be modelled as a continuation, compression, and renormalisation of rule refinement rather than the only stage at which provisional waking rules are adjudicated.

Status: architectural correction.

⸻

INV-COMMITMENT-001 — Action requires collapse

Any embodied agent acting in time must eventually collapse uncertainty into commitment. The system may preserve uncertainty internally, but behaviour requires selected action, inhibition, or delay.

Status: proposed invariant.

⸻

INV-UNCERTAINTY-001 — No waking rule is final

No rule apprehended or used during waking should be treated as final truth. Waking rule-use remains revisable through later evidence, exception handling, waking re-evaluation, and offline restructuring.

Status: proposed invariant.

⸻

Implementation sketch

Object: CandidateRule

CandidateRule rule_id domain proposed_regularities supporting_examples exception_examples N_observed e_exceptions tolerance_score context_tags affective_weight salience_weight goal_relevance approach_weight avoid_weight uncertainty confidence moral_residue_risk status: latent | rollout_eligible | action_tested | strengthened | weakened | split | merged | retired

⸻

Object: RolloutCandidate

RolloutCandidate rollout_id triggering_context active_candidate_rules predicted_future expected_reward expected_harm social_cost moral_cost energy_cost uncertainty approach_avoid_vector selected: true | false suppression_reason

⸻

Object: RuleActionEvidenceRecord

RuleActionEvidenceRecord event_id active_rules generated_rollouts selected_rollout suppressed_rollouts basal_ganglia_commitment action_taken immediate_outcome delayed_outcome prediction_error reward_signal harm_signal social_feedback moral_residue new_exceptions re_bucket_recommendations

⸻

Process sketch

Waking process

Observe event.
Retrieve similar episodes.
Detect candidate regularities.
Count or estimate exceptions.
Apply tolerance-like rule-availability threshold.
If threshold passes, mark candidate rule as rollout-eligible.
Hippocampal system generates candidate futures shaped by active rules.
Goal systems weight rollouts by approach/avoid/reward/threat/social/moral/energy factors.
Basal ganglia gates action, inhibition, sequence, or delay.
Evidence record is stored.
Waking systems continue to update rule confidence and rollout eligibility.

⸻

Sleep/offline process

Load evidence records from waking.
Identify rules repeatedly used in action selection.
Compare outcomes against predictions.
Detect exception clusters.
Split overbroad rules.
Merge redundant rules.
Renormalise affective weighting.
Update rollout eligibility thresholds.
Promote useful schemas.
Weaken or retire failed schemas.
Preserve moral residue markers where commitment under uncertainty caused harm or unresolved cost.

⸻

Clinical / psychopathology implications

This mechanism may help explain how maladaptive rules become powerful.

A rule may become clinically important not because the person explicitly believes it, but because it constrains what futures are available to imagine.

Examples:

Clinical pattern Possible REE interpretation Trauma schema threat rollouts become available too easily from high-salience evidence Paranoia social-threat rules bias hippocampal rollout generation toward hostile futures Depression negative self/world/future rules constrain imagined futures toward futility Obsessive-compulsive disorder danger/responsibility rules overgenerate catastrophic rollouts Anxiety avoidance rollouts become over-weighted relative to exploration or repair Mania reward/approach rollouts may be over-weighted and insufficiently checked by exception evidence Personality pathology attachment/social rules may strongly bias approach/avoid predictions in interpersonal contexts

This suggests a clinically useful phrasing:

The rule has become real because it controls the future you can imagine, not because you calmly decided it was true.

⸻

Moral residue implications

Moral residue arises cleanly at the commitment point.

The system may preserve uncertainty internally, but once basal ganglia gating commits a trajectory, other futures are excluded.

Moral residue may therefore attach to:

the selected action the suppressed alternatives the uncertainty at time of commitment the harm/benefit outcome the rule-field that biased the rollout the goal weights that pushed selection

This preserves the REE principle that moral cost is not only attached to outcomes, but also to action under uncertainty.

⸻

Open questions

Does the Tolerance Principle apply best to rule availability, rollout eligibility, action commitment, or all three at different levels?
How should REE represent exception counts where the evidence is affectively weighted rather than numerically counted?
Can a highly salient event distort tolerance thresholds enough to install a rule prematurely?
How does hippocampal replay during waking differ from sleep replay in refining candidate rules?
How should REE decide whether exceptions are noise, contextual modifiers, or evidence of hidden subcategories?
Should different domains have different tolerance thresholds?
Should harm/threat rules have lower thresholds for temporary installation but stronger requirements for promotion into durable schemas?
How should language-mediated rules interact with hippocampal rollout rules?
Can a verbal belief be weak while a rollout-shaping rule is strong?
How should REE detect when a rule is controlling imagined futures too narrowly?
Where should moral residue be stored: in the rule, the action, the selected rollout, or the whole evidence record?
How should sleep distinguish useful compression from pathological overgeneralisation?

⸻

Possible file destinations in REE_assembly

architecture/hippocampal_systems/rollout_selection.md architecture/control_plane/commitment_gating.md architecture/temporal_dynamics/waking_sleep_refinement.md architecture/social/convention_learning.md claims/mechanisms/tolerance_gated_rule_access.md claims/mechanisms/rules_as_rollout_selection_structures.md claims/invariants/action_requires_collapse.md

⸻

Compressed abstracted version

OBSERVATION: repeated_pattern + exceptions + context TOLERANCE_GATE: regularity becomes rule-available not final truth enough-to-use threshold HIPPOCAMPUS: generate candidate futures rules bias rollout eligibility exceptions alter rollout diversity GOAL_SYSTEMS: weight rollouts: approach | avoid | freeze | explore | repair | delay apply: reward | threat | attachment | energy | social cost | moral cost BASAL_GANGLIA: commitment bottleneck collapse weighted rollout field into: action | inhibition | sequence | delay EVIDENCE_RECORD: store active rules generated rollouts selected action suppressed alternatives outcomes prediction error harm/benefit moral residue WAKE_REFINEMENT: update rules during ongoing cognition SLEEP_REFINEMENT: replay | compress | renormalise | split | merge | prune | stabilise CORE: rules shape possible futures goals weight futures basal ganglia commits futures into action waking rule-use = experiment sleep refinement = continuation, not sole adjudication

⸻

Bottom line

This intake should revise the earlier model from:

basal ganglia learns rule → sleep refines rule

to:

distributed systems make candidate rules available → hippocampus embeds them into possible futures → goal systems weight those futures → basal ganglia commits one trajectory into action → evidence records support waking and sleep-based refinement

The Tolerance Principle is therefore most useful to REE as a candidate mechanism for rule availability and rollout eligibility, while basal ganglia gating remains the likely site where probabilistic rule-weighted futures are forced into embodied commitment.