Policy Primitive Granularity — Architectural Family

Status: family slot registered 2026-05-10 (ARC-069 parent + ARC-070 + ARC-071 candidate / pending_design). Family principle: the unit of policy that the rule-apprehension and diversity-generation layers operate on is itself a dynamic representation — primitives can decompose into finer ones (ARC-070, on prediction failure) or compose into coarser ones (ARC-071, on repeated grounding).

What this family fixes

REE today treats hippocampal rollout primitives as flat-grain. ARC-007 (strict) commits to value-flat hippocampal proposals, instantiated currently as CEM-sampled action sequences at single-timestep grain. Both downstream layers — ARC-062 rule apprehension and ARC-065 behavioural diversity generation — read these flat sequences. There is no machinery for the unit being chained to rescale.

The user’s observation, via the rollout chaining work:

“When looking at chaining behaviours together for hippocampal rollout we are working on there being behavioural diversity to chain together but when imagining a chain presumably the imagined chains may need splitting up into more chained behaviours or sometimes a number of chained behaviours together being realised to better be represented by a single behaviour (or rather than behaviour I suppose we should be thinking as policy or apprehended rule).”

This identifies two inverse architectural moves missing from the substrate:

Direction	Trigger	Timescale	Phenomenology
Decomposition (ARC-070)	Prediction failure on a chunk (low MECH-269 V_s on chunk’s region; or E2 disagreement; or completion-signal failure)	Fast, simulation-side	“I planned ‘go to the kitchen’ but now I need to plan the steps.”
Composition (ARC-071)	Repetition + outcome consistency on a sub-sequence	Slow, execution-side	“Making coffee was once seventeen motor acts; after a thousand repetitions it’s one unit.”

Both are operations on the same underlying object — the policy-primitive representation that ARC-007 emits, ARC-062 reads, and ARC-065 perturbs. ARC-069 names the parent commitment (“primitives are dynamic”); ARC-070 and ARC-071 name the inverse operations.

The three slots

ARC-069 — parent commitment (primitives are dynamic)

The architectural commitment, separate from any specific mechanism. Reads: the unit-of-policy that the rule-apprehension layer (ARC-062) and the diversity-generation layer (ARC-065) operate on is itself dynamic, not a fixed primitive. This is registered before mechanism work because the commitment cuts across multiple subsystems and constrains their joint design.

Without ARC-069, ARC-007 hippocampal proposals are flat-grain forever; ARC-062 must apprehend rules at whatever grain the proposer happens to emit; ARC-065 perturbs only at that grain. With ARC-069, the proposer emits at variable grain; ARC-062 sees rules at the appropriate scale for the current task; ARC-065 perturbs at the grain that actually adds diversity.

ARC-070 — decomposition on prediction failure (zoom in)

When an imagined chunk fails to ground at the resolution required for execution, the rollout representation re-segments the chunk into finer primitives. Trigger candidates:

MECH-269 V_s drop on the chunk’s region — cleanest existing-substrate hook; if V_s on the region is low, chunked primitives over that region can’t be reliably predicted.
E2 forward-model disagreement — chunk-grain prediction disagrees with stepwise primitive-grain prediction.
Hippocampal completion-signal failure (MECH-105 / ARC-028) — expected completion doesn’t arrive on schedule.
Mid-execution prediction error — chunk fired, agent is N steps in, next predicted state has diverged. Decompose the remainder.

Likely a combination — lit-pull will arbitrate.

Biology anchors (post-2026-05-10 lit-pull verdicts integrated; see evidence/literature/targeted_review_arc_070_decomposition/synthesis.md):

Zacks et al. 2007 (Psychol Bull) — event segmentation theory; PE is the canonical boundary trigger; framework is substrate-agnostic about whether the predictive stream is observed or imagined. R1 + R2 (theoretical) + R5 anchor. lit_conf 0.84.
Schacter, Addis & Buckner 2008 (Ann NY Acad Sci) — constructive episodic simulation; the same core network (medial PFC, hippocampus, retrosplenial / posterior cingulate, inferior parietal lobule) supports both remembering past and imagining future events. R2 LOAD-BEARING empirical anchor — added to the primary cluster from the lit-pull. lit_conf 0.82.
Badre & D’Esposito 2009 (Nat Rev Neurosci) — rostro-caudal prefrontal hierarchy; multi-level action abstraction. R3 empirical anchor. lit_conf 0.78.
Pfeiffer & Foster 2013 (Nature) — hippocampal forward path sweeps elaborate trajectories with cue-dependent dynamics; the imagination-side rollout substrate exists and is online during navigation. R1 substrate + R3 + R4 substrate anchor. lit_conf 0.78.
Koechlin & Summerfield 2007 (Trends Cogn Sci) — cascaded cognitive control; theoretical scaffold for multi-level decomposition. R3 theoretical + R5 hybrid anchor. lit_conf 0.74.
Schapiro et al. 2017 (Phil Trans Roy Soc B) — CLS within hippocampus; trisynaptic pattern-separation as candidate V_s substrate; dual pathways support multi-grain representation. R1 substrate + R2 pathway grounding + R3 multi-grain anchor. lit_conf 0.74.
McGovern & Barto 2001 (ICML) — bottleneck-state subgoal discovery; foil, not primary. R5 verdict distinguishes bottleneck-state framing (statistical) from PE-driven framing (biological); ARC-070 commits to PE-driven primary. lit_conf 0.62 (mixed direction).

ARC-071 — composition via repeated grounding (zoom out)

Sequences of primitives repeatedly executed together with consistent outcomes are accumulated into a single primitive. The chunked unit is added to the candidate pool; downstream layers operate on it atomically. The pre-existing primitive sequence remains available, so ARC-070 can decompose the chunk back if it later fails to predict.

ARC-071 names the transition mechanism that MECH-163 dual systems presupposes but does not specify. MECH-163 names the presence of a habit system and a planned system; ARC-071 is the machinery that pumps content from planned-into-habitual.

Trigger candidates:

Repetition count + outcome consistency — the canonical Graybiel pattern.
Reward-rate-conditioned chunking — Sakai 2003 motor sequence learning.
V_s-conditioned chunking — chunks form preferentially over sequences that produce reliable predictions (V_s-positive analog of ARC-070’s V_s-negative trigger).
Free-energy minimisation — chunks form because they reduce expected free energy.

Biology anchors:

Graybiel 1998 (Annu Rev Neurosci) + Graybiel 2008 — striatal chunking foundational references.
Sakai et al. 2003 (Curr Opin Neurobiol) — chunking in motor sequence learning across pre-SMA / cerebellum / striatum.
Wymbs et al. 2012 (Neuron) — motor chunk formation.
Sutton et al. 1999 (Artificial Intelligence) — options framework, cleanest formal analog.
Botvinick 2009 (Trends Cogn Sci) — hierarchical organisation of behaviour.
Yin & Knowlton 2006 (Nat Rev Neurosci) — DLS (habit / chunked) vs DMS (goal-directed / unchunked) division.
Smith & Graybiel 2013 (Neuron) — chunk dynamics: form, dissolve, reform across timescales.

How the three relate to existing REE substrate

ARC-007 strict (hippocampal proposals are value-flat): preserved. The regranularisation operates on the shape of proposals, not on their value. ARC-007’s no-internal-value-head invariant is intact.

ARC-062 rule apprehension (gated policy heads + context discriminator): reads policy primitives. With ARC-069 instantiated, the rules being apprehended are over potentially-chunked primitives. Open question deferred to child-MECH design: rule identification across grains.

ARC-065 behavioural diversity generation: perturbs primitives. Composition reduces local diversity (chunked sub-elements lose independent perturbability) but increases abstraction; decomposition increases local diversity but raises combinatorial cost.

MECH-269 V_s (per-stream / per-region verisimilitude): the most plausible trigger signal for ARC-070 decomposition, and a candidate condition (V_s-positive) for ARC-071 composition. ARC-070 becomes the second major V_s consumer after MECH-269b symmetric-V_s-gating.

MECH-288 event segmenter (substrate-side two-level hierarchical detector reading latent + PE streams): ARC-070 is a bidirectional consumer of the same substrate, not a parallel module (lit-pull R2 verdict 2026-05-10, LOAD-BEARING, conf 0.74). MECH-288’s API consumes latent_dict + pe_dict — substrate-agnostic about whether those latents originate in observation or in simulation. ARC-070 adds the policy-side consumer that reads MECH-288 boundary pulses on the rollout / imagination stream and re-segments the proposal trajectory at finer grain. Same detector, two input streams (observation, rollout), two downstream effects (MECH-269 anchor-set update on observation; decomposition pulse on rollout). The “imagination-side analog” framing from the original registration is superseded by the bidirectional-substrate framing. Falsifiable: lesioning MECH-288 should impair BOTH observation-side anchor partition AND imagination-side decomposition; the bidirectional hypothesis predicts no dissociation. See evidence/literature/targeted_review_arc_070_decomposition/synthesis.md for the constructive-episodic-simulation grounding (Schacter 2008 core network).

MECH-292 / MECH-293 ghost-goal bank + waking probes: anchors are point-keys. ARC-069 + ARC-071 add structure-bearing primitives that anchors can seed. The SD-039 goal_payload extends naturally to chunks — a chunked primitive with a stable goal-payload is an accumulated “recipe” the agent has formed.

MECH-163 dual systems (planned vs habitual): ARC-071 is the transition mechanism. Habit chunks are the output of ARC-071; planned-system trajectories are the input. Without ARC-071 the dual systems are static configurations; with ARC-071 the division of labour shifts continuously with experience.

MECH-094 hypothesis_tag: ARC-070 fires during simulation (must respect MECH-094 — decomposition during waking action does not write residue). ARC-071 fires from real executed sequences (hypothesis_tag=False) by default, with one explicit narrow exception: MECH-322 (registered 2026-05-11 to resolve the ARC-071 lit-pull R6 governance escalation) permits ARC-071 chunking from replayed (hypothesis_tag=True) sequences during designated SD-017 sleep phase IF (a) the replayed sequence carries a value-tag from prior real-executed episodes meeting a high-positive threshold, (b) the replay occurs in sleep mode (not waking DMN), (c) the formed chunk carries a replay_origin=True audit flag and faces accelerated dissolution if not corroborated by real waking execution within N episodes (suggested default 50–100). MECH-094 strict gating remains the default for all other write paths; MECH-322 is a parallel narrow path with audit trail. Waking DMN (where MECH-292 / MECH-293 ghost-goal probes operate) remains MECH-094-strict — the carve-out is sleep-only. The carve-out matches the Albouy 2013 hippocampal-striatal sleep-replay coupling found in the ARC-071 lit-pull R6 evidence.

How the three slots compose at runtime

Hippocampal rollout produces a candidate trajectory.
                         |
                         v
Trajectory consists of a sequence of *primitives*. Each primitive is
EITHER a single-action unit OR a chunk (ARC-071-formed, ARC-070-decomposable).
                         |
                         v
For each primitive in the trajectory, evaluate predictability:
  - chunk grain: V_s on chunk region (MECH-269) + chunk outcome confidence
  - if V_s is below threshold, decompose chunk into sub-elements (ARC-070)
                         |
                         v
ARC-062 apprehends rules over the (possibly-decomposed) trajectory.
ARC-065 generates diversity at the appropriate grain.
                         |
                         v
On commit + execution, accumulate (sub-sequence, outcome) pairs.
                         |
                         v
ARC-071 composition accumulator: for repeated (sub-sequence -> outcome)
patterns with low outcome variance, form a new chunked primitive and add it
to the candidate pool. (Original primitives remain; chunks are additive.)

ARC-070 sits in the rollout path (simulation-side, fast); ARC-071 sits in the post-execution path (real-grounding-side, slow). They operate on the same representational object but at different stages and timescales.

Falsifiable family-level prediction

Run an REE agent for many episodes on a task with a structurally repeating sub-sequence (e.g. SD-054 reef + a repeating forage-loop):

Without ARC-071: every rollout in episode 1000 still proposes at single-action grain. Rollout deliberation cost stays flat. Behavioural latency on the repeating sub-sequence does not drop.
With ARC-071 active: the repeating sub-sequence compresses into a chunked primitive. Rollout deliberation cost on it drops measurably. Behavioural latency drops (faster commit, less deliberation).

Then introduce a region of high V_s drop in the middle of a chunked primitive’s region:

Without ARC-070: agent commits to the chunk, executes it blind, pays prediction-failure cost at execution.
With ARC-070 active: agent re-decomposes during rollout; refines the proposal at finer grain over the low-V_s region; either finds a finer-grain plan that grounds, or fails-to-plan and aborts (both PASS — the prediction is the decomposition fires, not that the alternative succeeds).

What this commits REE to

A representational shape for policy primitives that has not previously been registered: variable-grain, dynamically restructured by both prediction failure (decompose) and repetition with consistency (compose). Every existing slot in REE that touches the rollout or rule-apprehension paths has been written assuming flat-grain primitives; ARC-069 commits to relaxing that assumption.

The biggest downstream consequence is for ARC-062 / ARC-065. Rules apprehended at one grain may not apply at another; diversity generated at one grain may not produce useful candidates at another. The grain question is now part of the rule and diversity design space — not a parameter we set, but a state-conditioned variable. This is bigger than it looks; the lit-pulls and child-MECH phases will need to address grain-invariant rule apprehension explicitly.

What this is NOT

Not a hierarchical-RL options re-implementation. Sutton 1999 is the analog, not the spec. Biology distributes the substrate (hippocampal / prefrontal / striatal) in a way that doesn’t cleanly map to the options framework’s value-side machinery.
Not a representation-compression slot at the encoder layer. Chunks are over policy primitives, not over latent observations. SD-009 z_world contrastive supervision and SD-018 z_world resource proximity are encoder-side; ARC-069 family is policy-side.
Not a goal hierarchy. Chunks are over policy primitives. Goal hierarchies are a separate architectural concern (cf. SD-015 z_resource and z_goal_snapshot in MECH-292/293). The two interact (goals drive which chunks form) but are not the same slot.
Not a feedback loop on V_s itself. ARC-070 reads V_s; the V_s update path remains MECH-269 substrate.

Open Q-claims worth registering at child-MECH time

Grain-invariant rule apprehension: how does ARC-062 handle rules learned at one grain that need to apply at another? Hierarchical rules are likely apprehended at multiple grains in parallel (Botvinick 2009); the design will need to address this.
Chunks-of-chunks recursion: depth of recursion permitted. Biology supports deep hierarchy; computational cost / catastrophic-rigidity tradeoffs are open.
Decomposition triggered by outcome inconsistency: when an existing chunk’s outcome variance rises above threshold, does the chunk decompose (ARC-070), or simply become un-selectable from the proposal pool? Different behavioural signatures; lit not decisive.
Diversity generation across grains: should ARC-065 operate on chunked or unchunked representations? Likely both at different timescales — separate Q-claim.

Pathology cross-references (speculative, pending lit-pull)

These cross-link to psychiatric_failure_modes.md after lit-pulls land:

Obsessive-compulsive disorder as ARC-071 weakened: chunks fail to form, every routine action is re-planned at primitive grain, over-deliberation signature.
Autism / insistence-on-sameness as ARC-070 weakened on chunked primitives: chunks form normally but cannot be re-decomposed when context demands different grain, producing rigidity.
Skill / motor-learning impairments: chunking machinery itself impaired; sequences never compress; deliberation budget exhausted.

These are pre-registration cross-references — speculative until biology lit-pulls land and child-MECH mechanism shape is fixed.

Child MECH/SD design — what’s needed before phasing

Two independent lit-pulls anticipated:

targeted_review_arc_070_decomposition/ — LANDED 2026-05-10. 7 entries; aggregate lit_conf 0.88 (indexer-computed); 6 supports + 1 mixed. Five verdicts settled: R1 V_s-drop on chunk’s region as primary trigger (PE-driven); R2 LOAD-BEARING — SHARED SUBSTRATE: ARC-070 is a bidirectional consumer of MECH-288, not a parallel module; R3 multi-level recursive decomposition with depth cap 3-4; R4 both pre-commit and mid-execution phases via same mechanism with MECH-094 hypothesis_tag-conditional downstream effects; R5 PE-driven primary, McGovern-Barto bottleneck-state framing as optional consolidation-phase secondary. See evidence/literature/targeted_review_arc_070_decomposition/synthesis.md.
targeted_review_arc_071_composition/ — LANDED 2026-05-10 (sibling parallel pull). 9 entries; aggregate lit_conf 0.848. R3 confirmed ARC-071 IS the missing transition mechanism MECH-163 dual_goal_directed_systems presupposes (MECH-163 depends_on +ARC-071 committed 2026-05-10). R6 SAFETY-CRITICAL: biology does NOT cleanly gate chunking write path against replay/imagined sequences — RESOLVED 2026-05-11 via MECH-322 sleep-replay value-conditioned carve-out (narrow exception path, audit trail, accelerated dissolution on uncorroborated replay-origin chunks). Child-MECH design unblocked.

The two children share the parent commitment but no execution path — they’re decoupled within the cluster, similar to how ARC-066 / ARC-067 / ARC-068 are decoupled within the non-deficit-action-drives family.

Child-MECH design — ARC-070 side (REGISTERED 2026-05-10)

With R2 settled, ARC-070’s first child mechanism is registered as MECH-321 policy.decomposition_via_event_segmenter (candidate / v3_pending):

Subject: policy.decomposition_via_event_segmenter. Policy-side consumer of MECH-288 boundary pulses on the rollout / imagination input stream.
depends_on: ARC-070 (parent), MECH-288 (substrate), MECH-269 (V_s primitive trigger source), MECH-094 (hypothesis_tag gating).
Trigger (R1): V_s drop on chunk’s region (read out of MECH-288 + MECH-269). MECH-288 produces the boundary; MECH-321 consumes the boundary at the policy-primitive layer and triggers re-segmentation.
Output: re-segmented rollout proposal stream at finer grain.
Depth control (R3): recursive multi-level decomposition with depth cap 3-4.
Phase handling (R4): pre-commitment fires during simulation under hypothesis_tag=True with no residue write; mid-execution decomposition fires under hypothesis_tag=False with residue write enabled and observation-side consumer chain operating normally.
Optional secondary (R5): bottleneck-aware consolidation-phase analysis; may integrate with ARC-071 chunk-formation pipeline once the ARC-071 R6 governance decision lands.

Substrate-readiness prerequisite: MECH-288 substrate (event_segmenter.py) must land first WITH the input_stream label extension (per MECH-288 2026-05-10 ARC-070 bidirectional-consumer commitment in its notes). MECH-321 cannot be substrate-implemented before MECH-288. Phase-1 placement is a thin policy-side module subscribing to a MECH-288 BoundaryEvent queue filtered to input_stream=rollout, wired at the hippocampal-rollout candidate-generation layer prior to E3 trajectory selection.

Discriminative-pair validation experiment specified in MECH-321 functional_restatement (ARM_0 baseline / ARM_1 V_s-drop primary / ARM_2 bottleneck-state primary); deferred until substrate lands.

Child-MECH design — ARC-071 side (BOTH OPERATORS REGISTERED 2026-05-11)

With R6 resolved via MECH-322 (sleep-replay carve-out), ARC-071’s child-MECH design is unblocked. The lit-pull’s R2 verdict (phase-dependent multi-substrate with formation in striatum/DLS and maintenance in IL/vmPFC) maps onto two child-MECHs, mirroring Smith & Graybiel 2013’s “dual operator view”:

MECH-323 policy.composition.chunk_accumulator_formation (REGISTERED 2026-05-11) — the striatum/DLS-analog formation operator.

Subject: policy.composition.chunk_accumulator_formation. Builds chunk candidates from sequences of policy primitives repeatedly executed together with consistent outcomes.
depends_on: ARC-071 (parent), MECH-094 (default strict gate), MECH-322 (sleep-replay carve-out path), SD-014 (valence vector), SD-039 (anchor goal-payload), MECH-269 (V_s positive secondary).
Trigger conditions (joint AND): (1) repetition count ≥ R_min over sliding window W; (2) outcome-variance below F_low formation threshold (hysteresis with MECH-324’s F_high dissolution threshold); (3) evaluative gate — accumulated outcome mean must be positive (Graybiel 2008 framing). Secondary preference: V_s-positive (the symmetric inverse of MECH-321’s V_s-negative trigger).
Chunked-primitive object fields: sequence, initiation_set, termination_condition, value_tag, replay_origin, formation_timestamp, depth. The initiation_set + termination_condition fields satisfy the Sutton 1999 options-structure requirement that R4 surfaced.
Suggested parameter defaults (child-MECH validation refines): R_min = 20 reps; W = 100 trials; F_low = 0.15 (on 0–1 normalised outcome-variance); evaluative threshold = baseline + 0.05; recursion depth cap = 3.
Two write paths: default MECH-094-strict (real-executed sequences, replay_origin=False); MECH-322 sleep-replay carve-out (high-value-tagged sleep replays, replay_origin=True, dissolution-deadline accounting).
What MECH-323 preserves: ARC-007 strict value-flat proposals (value_tag is metadata, not a value head); MECH-094 strict on the default path; SD-039 anchor payload semantics; MECH-269 substrate.
First validation experiment: substrate-readiness diagnostic measuring whether the accumulator fires at all on a structurally repeating sub-sequence task with default parameters. Behavioural-latency / rollout-cost measurements follow once the accumulator-only diagnostic passes.

MECH-324 policy.composition.chunk_maintenance (REGISTERED 2026-05-11) — the IL/vmPFC-analog maintenance operator. Causally required for chunk crystallisation per Smith & Graybiel 2013 selective IL optogenetic disruption (lit_conf 0.86).

Subject: policy.composition.chunk_maintenance_dissolution. Operates on MECH-323-formed chunks; gates whether they crystallise into selectable proposals and how long they persist.
depends_on: ARC-071 (parent), MECH-323 (the formation operator MECH-324 maintains), MECH-322 (sleep-replay carve-out; accelerated-dissolution rule lives here), MECH-163 (dual-systems substrate; MECH-324 is habit-maintenance side), INV-037 (vmPFC stored/active distinction), INV-038 (vmPFC EVR pattern).
Four-state lifecycle: FORMING (chunk just created by MECH-323, weak selection_weight, crystallisation_counter accumulates on successful real executions) → CRYSTALLISED (counter ≥ C_min; full selection_weight; persists until F_high variance or MECH-322 deadline) → DISSOLVING (variance > F_high; selection_weight decays linearly over T_dissolve trials; recovery possible if variance drops back) → DISSOLVED (removed from proposal pool; retained in audit-trail register). MECH-322 replay-origin uncorroborated chunks transition DIRECTLY from CRYSTALLISED to DISSOLVED on deadline, bypassing the slower DISSOLVING window.
Three sub-mechanisms (component decomposition, not separate child-MECHs): (A) crystallisation counter — real-execution-only; replay/hypothesis_tag=True do not increment; (B) outcome-variance dissolution gate — sliding-window variance over W_maint trials, threshold F_high, slow T_dissolve timescale per R5; (C) replay-origin accelerated-dissolution gate — implements the MECH-322 corroboration-deadline rule.
Suggested parameter defaults (child-MECH validation refines): C_min = 5 corroborating executions; W_maint = 100 trials; F_high = 0.45 (hysteresis gap with MECH-323’s F_low = 0.15); T_dissolve = 50 trials; N = 75 waking episodes (replay-origin corroboration window).
Substrate hooks: crystallisation_state via INV-037 stored/active distinction (stored = state field on chunk object; active = selection_weight at current step); maintenance decisions via INV-038 EVR pattern (running outcome valence over post-formation executions feeds F_high gate).
What MECH-324 preserves: ARC-007 strict value-flat proposals (selection_weight is a candidate-machinery property, not a value head); MECH-094 strict default; MECH-322 carve-out semantics; MECH-323 formation semantics.
Falsifiable architectural prediction: an REE agent with MECH-323 active but MECH-324 disabled should produce chunks that form correctly but never crystallise (accumulator fires, behavioural latency does NOT drop). MECH-324’s contribution is operationally the latency drop / rollout-cost reduction signature, mirroring Smith & Graybiel 2013’s IL-disruption-vs-intact contrast.
First validation experiment design: three-arm discriminative test — ARM_0 baseline (no chunking), ARM_1 MECH-323 only (formation without maintenance), ARM_2 MECH-323 + MECH-324 (full machinery). The ARM_1 vs ARM_2 contrast isolates MECH-324’s contribution.

ARC-071 substrate is now fully specified. Formation (MECH-323) + maintenance (MECH-324) + sleep-replay carve-out (MECH-322) form the complete chunking machinery. Next architectural pass: substrate-level implementation in ree_core via the /implement-substrate skill, followed by the three-arm validation experiment.

Frontoparietal early-phase parsing (Wymbs 2012 R2 partial-mapping): not a separate ARC-071 child-MECH. It bleeds into ARC-070 / MECH-321 territory (segmentation, not concatenation); cross-linked to MECH-321 rather than registered here.