Waking-Phase V_s Invalidation: Runtime Trigger + Local Accumulator

Claim cluster: MECH-287 (broadcast invalidation trigger, NEW), MECH-284 (V_s residual accumulator, REFINED operational definition), MECH-269 (anchor selection + dual-trace reset criteria, REFINED) Status: candidate, v3_pending Phase status (substrate): Phase 1 (per-stream V_s) IMPLEMENTED; Phase 2 (event segmenter MECH-288, anchor sets, per-region V_s, invalidation trigger MECH-287) IMPLEMENTED; Phase 3 online arm (MECH-284 staleness accumulator + MECH-269 hysteresis consumer) IMPLEMENTED 2026-04-24; Phase 3 offline arm (MECH-285 sleep-priority readout) DEFERRED. Registered: 2026-04-22 Origin exemplar: V3-EXQ-475 (re-run of V3-EXQ-471 with SD-036 GABA decay enabled) showed 5-6 PAG freeze releases per seed but ~12 re-commits per release, with all 1000/1000 eval steps in freeze. SD-036 decay and MECH-279 freeze exit are firing correctly; the endogenous driver of re-commit is the hippocampal proposer pulling avoid-flavoured trajectories from the original anchor. There is no waking-phase signal that downweights the anchor as evidence accumulates that it no longer fits. Lit-pulls:

  • evidence/literature/targeted_review_waking_v_s_invalidation/ (12 papers + SYNTHESIS.md, 2026-04-22) — origin pull, drove MECH-287 registration.
  • evidence/literature/targeted_review_v_s_foundation/ (12 papers + SYNTHESIS.md, 2026-04-22) — foundation pull commissioned for Phase 2 substrate work; covers schema-region granularity (verdict 1), multi-map coexistence and dual-trace modes (verdict 2), per-stream V_s as projection-readout of mixed-selectivity integrated code (verdict 3), splitter-cell anchor types (verdict 4), and the upstream anchor-side comparator stage of the MECH-287 trigger (verdict 5).
  • Cross-references: targeted_review_connectome_mech_269/ (replay content), targeted_review_homeostatic_override/ (LC/orexin coupling). Depends on: SD-005, ARC-007, ARC-018, MECH-089, MECH-094, MECH-269, MECH-272, MECH-285, SD-010, SD-011

Problem

The V_s bidirectional cluster registered 2026-04-22 (MECH-283 forward recognition gate, MECH-284 residual schema-staleness accumulator, MECH-285 sleep replay priority) closes the offline arm: residuals accumulated across waking get integrated and acted on during sleep. The runtime arm is incomplete. MECH-284 specifies that staleness accumulates and is “read-only during waking; consumed during sleep by MECH-285.” This works for the slow schema-revision regime but cannot rescue an agent that is currently trapped in an avoid mode driven by a stale anchor.

V3-EXQ-475 made this concrete. Decay returns z_harm_a below the freeze exit threshold; the freeze gate releases (5-6 times across the 1000-step eval). But the hippocampal proposer keeps drawing trajectories from the original avoid-anchor; those trajectories re-stoke z_harm_a; freeze re-commits within ~80 steps. The lock is not in the harm stream itself (decay works) and not in the freeze gate (exits work). The lock is in the proposer’s choice of anchor, which has no waking-phase mechanism for invalidation.

The biological expectation is that anchor invalidation happens during waking too – otherwise an agent could not correct a stale schema across a single afternoon, and clinical phenomenology would not include experiences like “I just realised the threat is gone” in real time. The architectural commitment MECH-287 + the refinements to MECH-284 and MECH-269 make is that:

  1. Waking-phase invalidation is a two-circuit system. A trigger event signals single-instance schema-fit violations broadly; a local accumulator integrates these over a per-schema window and emits a downweight on the affected anchor’s V_s.
  2. The two circuits are substrate-distinct. The biology supports separating them (LC/DA for the broadcast trigger; OFC for the local accumulator), and the failure modes dissociate clinically.
  3. Reset is dual-trace. When V_s drops below the anchor-reset threshold, the predecessor anchor is marked inactive, not erased. The new anchor is added; routing (MECH-272) decides which is operative under what state.

Foundation lit-pull integration (2026-04-22 Phase 2)

The targeted_review_v_s_foundation/ SYNTHESIS (12 papers; place-cell remapping, attractor dynamics, splitter/trajectory cells, schema cells / event boundaries, multi-trace / pattern separation, comparator / HC-VTA loop, subicular routing) constrains four substrate decisions in addition to the MECH-287 dual-component refinement (verdict 5, integrated into Section 1 below).

Verdict 1 — schema-region granularity (NOT place-cell-field). The substrate defaults the V_s region unit to schema / event-segment / action-object scale, with multi-scale support where the task demands finer resolution. Eichenbaum 2017 establishes that hippocampal coding is multi-dimensional mixed-selectivity at the cell level with the relevant organisation scale emerging from task structure; Tse 2007 schemas operate at integrated event/context level; Sols/DuBrow/Davachi 2017 shows event boundaries (not place transitions) drive reinstatement and consolidation priority. Substrate implication: attribution_weight(r, source_streams) operates over schema regions, not place fields. The accumulator staleness[r] is keyed on schema-region IDs. Place-cell-field-scale resolution should be available as an option but is not the default unit. Region granularity itself is also a tunable lever in scaling experiments — see project memory project_intelligence_scaling_levers.md.

Verdict 2 — dual-trace anchor preservation supports both hard-switching AND soft re-weighting. Wills 2005 (attractor morphing, bistable winner-take-all ensemble switching) gives the hard-switch mechanism; Colgin 2008 (rate remapping) and Frank 2000 / Wood 2000 (trajectory-cell rate modulation) give the soft-re-weighting mechanism. Substrate implication: MECH-269 dual-trace must support BOTH a hard anchor switch (Wills attractor flip) and a soft mixture re-weighting (Colgin-style graded mixture). MECH-272 routing then treats the choice between these regimes as a function of V_s magnitude and input ambiguity rather than committing architecturally to one mode. Section 3 below (mark_inactive(a) + seek_new_anchor) describes the hard-switch path; the soft-re-weighting path is added as a separate routing-level operation that adjusts mixture weights over the active anchor set without marking any anchor inactive.

Verdict 3 — per-stream V_s is a projection-readout of an integrated mixed-selectivity code, NOT a biological per-stream computation. Eichenbaum 2017: hippocampal CA1 carries multiplexed independent representations of multiple dimensions with mixed selectivity at the single-cell level — biology computes a single integrated mixed code; per-stream readout is downstream extraction. Vinogradova 2001 grounds the principle that biology computes mismatch between functionally distinct streams (the principle is biological; the specific five-stream count z_world / z_self / z_harm_s / z_harm_a / z_goal is an engineering choice). Yonelinas 2019 contextual-binding theory shows the integrated code degrades into something more per-stream-like under capacity constraints, giving the substrate plan a recoverable-fallback grounding. Substrate implication: register per-stream V_s as a projection-readout in the substrate documentation and naming (v_s_per_stream_readout, not v_s_per_stream_compute). This validates the Phase 1 deviation from the original plan that wired V_s uniformly via an identity proxy at HippocampalModule scope rather than per-stream forward predictors — architecturally that is the more biology-faithful choice, since the integrated code lives at the hippocampal scope and per-stream readouts are extracted from it. Per-stream forward predictors should NOT be added in Phase 2.

Verdict 4 — splitter-cell support for both pure-place and stream-conditioned anchors. Wood 2000 (~one-third pure-place, ~one-third context-dependent firing on shared stem in continuous-alternation) and Frank 2000 (W-maze trajectory-cell encoding) establish that anchor encoding conditional on a stream-mixture / trajectory / context signal is biologically standard, not exotic. Substrate implication: MECH-269 anchor type catalogue includes BOTH pure-place anchors (V_s reset on the location signal alone) and stream-mixture- conditioned anchors (V_s reset conditional on a trajectory / context / intent signal). The exact ratio is task-dependent and is NOT hard-coded; both types must be supported and the mix emerges from training.

Mechanism

1. MECH-287 (NEW): Dual-component invalidation trigger

A per-step broadcast signal that fires whenever a schema-fit violation is registered above threshold. Computationally:

trigger(t) = { norm(z_pred(t-1) - z_obs(t)) > theta_trigger      OR
               value_PE(t) < -theta_value                          OR
               external_surprise_signal(t) > theta_surprise }

When trigger(t) fires, an event is emitted on a global trigger bus carrying:

  • t (time index)
  • magnitude (which gate fired and by how much)
  • source_streams (which streams contributed – z_world, z_harm, z_goal, etc.)
  • current_anchor_set (which anchor regions were active when the violation occurred)

The trigger event is the thing the accumulator integrates. It is not itself an update – it is a labelled event broadcast widely so that all downstream accumulators that care about this region can consume it.

Biological substrate — DUAL-COMPONENT coupled-stage architecture (per targeted_review_v_s_foundation/ SYNTHESIS verdict 5; supersedes the broadcast-only framing in the original 2026-04-22 design):

The trigger is not a single broadcast event but the downstream stage of a two-stage loop. Treating broadcast and comparator as alternative trigger sources misses the loop architecture biology actually uses.

Upstream anchor-side comparator stage:

  • CA1/CA3 mismatch comparator (Vinogradova 2001) — hippocampus computes match/mismatch between two functionally distinct input streams; output gates both intra-hippocampal output and subcortical arousal modulation.
  • Subicular routing (O’Mara 2009) — anatomical hub through which hippocampal mismatch reaches NAc-VP-VTA. The trigger flows through this pathway; it does not bypass it.
  • HC-VTA loop wiring (Lisman & Grace 2005) — hippocampal mismatch propagates via subiculum-NAc-VP to VTA, which back-projects dopamine to enhance hippocampal LTP. This loop is what couples the upstream comparator to the downstream broadcast.
  • Event-boundary reinstatement support (Sols/DuBrow/Davachi 2017) — event boundaries behave like local mismatch signals at the schema/event-segment scale, consistent with the comparator firing on schema-region boundaries rather than arbitrary single-step latent deltas.

Downstream broadcast stage (parallel, partially redundant):

  • Locus coeruleus phasic burst (Aston-Jones & Cohen 2005; Sara & Bouret 2012)
  • DA generalised PE (Schultz 1997; Gardner, Schoenbaum & Gershman 2018) — value-PE arm
  • Lateral habenula negative-RPE (Matsumoto & Hikosaka 2007; Bromberg-Martin & Hikosaka 2011) — value_PE < -theta_value arm

Dissociation inside MECH-287: comparator-loss = silent trigger (CA1/CA3 lesion analog: no upstream signal to broadcast); broadcast-loss = silent invalidation (LC lesion analog: comparator fires but no global gain modulation downstream); both intact = full trigger event. This dissociation is testable separately from the trigger-vs-accumulator (MECH-287 vs MECH-284) factorial.

2. MECH-284 REFINED: Local schema-staleness accumulator

The MECH-284 entry currently says staleness is “read-only during waking; consumed during sleep by MECH-285.” This is correct for the schema-revision read-out, but the accumulator itself must update during waking — that is the only opportunity to integrate trigger events into a per-region signal. The clarification is:

For each schema region r in active_anchor_set(t):
    if trigger(t):
        staleness[r] += attribution_weight(r, source_streams) * magnitude
    staleness[r] *= leak_factor   # slow decay so stale regions clear

attribution_weight(r, source_streams) is the schema-specific credit assignment step: which schema regions plausibly contributed to the broadcast event. Biologically this is the cortical control over LC (ACC/OFC -> LC) — the broadcast trigger is undifferentiated; the local accumulator does the credit assignment.

The accumulator is read by two downstream consumers:

  1. MECH-285 (offline / sleep): the staleness map biases replay priority during sleep consolidation. This is the existing read-out.
  2. MECH-269 anchor-reset gate (online / waking): when staleness[r] exceeds theta_reset(r), the anchor on region r is marked inactive and a new anchor is sought for the current latent slice. This is the new read-out specified here.

The two read-outs share the same accumulator state but operate on different timescales — sleep consumes the integrated map across many anchors; the online reset consumes a single anchor’s staleness against its own threshold.

Biological substrate: OFC representational drift (Wilson et al 2014; Stalnaker et al 2015) is the best candidate for the local accumulator state. The OFC-as-cognitive-map-of- task-space framing maps directly onto V_s as per-region schema validity; OFC labelling integrates evidence over trials and reorganises when the underlying state is judged to have changed. Hippocampal DG / alEC pattern separation (Yassa & Stark 2011; Reagh et al 2018) provides upstream interference-detection signals that feed the accumulator.

3. MECH-269 REFINED: Anchor-reset criteria with dual-trace preservation

MECH-269 currently specifies anchor selection by V_s gating. It does not specify what happens when V_s drops below threshold for an active anchor. The refinement:

For each active anchor a on region r:
    V_s_anchor(a, t) = V_s(r, t) - staleness[r]
    if V_s_anchor(a, t) < theta_reset(r) for k consecutive steps:
        mark_inactive(a)
        seek_new_anchor(latent_slice(t))

Dual-trace requirement (Bouton 2004). mark_inactive(a) must NOT erase the anchor. Extinction is new learning, not unlearning; the avoid-anchor and a new safety-anchor must coexist in the hippocampal anchor set, with MECH-272 (state-gated routing) deciding which is operative under what state. This makes MECH-272 a load-bearing part of the fix, not just an ancillary mechanism. State-gated routing has to handle anchor mixtures in waking under the same logic it uses for waking-anchor vs sleep-probe routing.

k consecutive steps requirement. A single trigger event should not collapse an anchor — the anchor would oscillate. Requiring k consecutive steps below threshold gives the architecture a hysteresis property analogous to MECH-266 asymmetric mode hysteresis. Initial value k = 5 (substrate-tunable).

Why threshold not graded. The accumulator is graded; the reset is threshold. Mixing the two avoids a discrete reset on every staleness perturbation while keeping the graded accumulator informative for offline replay priority.

Why these three mechanisms together rather than expanding MECH-284 alone

The minimal-change alternative would be: keep MECH-284 alone, expand its functional restatement to include the trigger event implicitly, expand MECH-269 to include anchor reset by reading MECH-284 output. This is cheaper to register but gives up two things:

  1. Failure-mode dissociation. LC hypoactivity (under-trigger), DA dysfunction (under-value-trigger), LHb hyperactivity (over-trigger), OFC lesion (accumulator broken), alEC hypoactivity (input pathway broken) are biologically distinguishable failures that produce distinguishable phenotypes (perseveration, anhedonia, learned helplessness, reversal-learning failure, mnemonic discrimination loss). A single MECH-284 covering both the trigger and the accumulator collapses this distinction.

  2. Substrate-readiness story. A trigger-event MECH can be implemented in ree-v3 as a per-step signal in the agent loop (similar to how prediction-error signals are already wired). The accumulator MECH-284 needs a longer-timescale state machine. Separating them lets each be implemented and tested independently.

The cost of registering MECH-287 separately is a single new claim entry. The benefit is an architecture that maps cleanly onto biology and onto the experiments that would falsify each component.

Predicted observables (V3 scope)

A V3 experiment validating MECH-287 + refined MECH-284 + refined MECH-269 would measure:

  1. EXQ-475 freeze re-commit resolution. Re-running V3-EXQ-475 with the runtime invalidation circuit enabled (default theta_trigger, theta_reset, k=5) should produce: trigger events fire on each freeze release as the agent observes that no harm follows; staleness[r] accumulates over ~3-5 trigger events; anchor reset fires; new anchor selected (probably nearer to current safe location); freeze re-commit rate falls dramatically. Predicted: 1-2 freeze episodes total instead of 20+, eval ends with the agent exploring rather than locked.

  2. Trigger-vs-accumulator dissociation. Lesion the accumulator only (set attribution_weight to zero, so trigger events fire but never accumulate) — agent should look like vanilla EXQ-475 (releases occur, re-commits dominate). Lesion the trigger only (set theta_trigger very high, accumulator can never integrate) — agent should look the same. Lesion both — same. Lesion neither — resolution. The four-arm factorial is the falsifiable architectural prediction.

  3. Dual-trace preservation. After the agent escapes the avoid-mode, present the original threat cue again. The avoid-anchor should still be retrievable (response latency, anchor-set inspection) — i.e. extinction does not erase. If MECH-272 routing is intact, the agent should respond to the cue again rather than ignoring it.

  4. Sleep replay priority coupling (Q4 from SYNTHESIS). A waking session with high trigger-event load should produce higher MECH-285 replay priority on the affected regions than a session with low trigger-event load, even when the offline schema inputs are matched. This is the architectural coupling that the lit-pull identified as underdetermined; the V3 experiment can test it directly.

Candidate experiment names (not yet queued):

  • v3_exq_NNN_mech287_runtime_invalidation_unlocks_exq475.py (single-condition: full circuit on/off, replicate EXQ-475 conditions; default candidate for V3-EXQ-476)
  • v3_exq_NNN_trigger_vs_accumulator_dissociation.py (four-arm factorial)
  • v3_exq_NNN_dual_trace_post_extinction_retrieval.py (extinction + re-cue)
  • v3_exq_NNN_waking_load_to_sleep_priority.py (Q4 coupling)

Substrate hooks required:

  • ree_core/regulators/invalidation_trigger.py — new module hosting MECH-287; emits trigger events on a queue consumed by the accumulator.
  • ree_core/hippocampal/staleness_accumulator.py — new module hosting MECH-284 runtime state; subscribes to the trigger queue, applies attribution_weight, maintains the per-region staleness map; exposes both the sleep read-out (existing MECH-285 consumer) and the online reset read-out (new MECH-269 consumer).
  • ree_core/hippocampal/module.py — extend HippocampalModule with anchor-reset gate reading the staleness accumulator; preserve old anchors via dual-trace inactive flag.
  • ree_core/cingulate/salience_coordinator.py — extend MECH-272 routing to handle anchor mixtures (active + inactive coexist; routing decides which is operative).

MECH-285 offline arm — design (informed by 2026-04-24 lit-pull)

The sleep-priority readout that consumes the MECH-284 staleness map. Implementation deferred from the Phase 3 substrate commit; design commitments below are informed by evidence/literature/targeted_review_mech285_sleep_replay_seed/SYNTHESIS.md (10 papers, five architectural verdicts).

1. Seed coverage: BROAD-UNBOUNDED, experience-structured. The start-state distribution for sleep-phase replay should be drawn from the union of the current active anchor set AND recently-inactive anchors (including dual-trace-inactive anchors created by MECH-269 extinction events), weighted by staleness. Not a narrow active-only pool. Evidence: Karlsson & Frank 2009 (awake replay routinely covers remote environments not part of the current active set); Gillespie/Liu 2021 (replay enriched for past-rewarded AND not-recently-visited locations); Olafsdottir 2015 (preplay extends beyond visited space toward goal-biased unvisited regions). No published time-window cutoff — the lit-pull found no evidence that the sampler needs a bounded replay window, so the substrate should be stateless on “how far back to look.”

2. Priority shape: STALENESS-PROPORTIONAL, not threshold-gated. Replay-event count per region scales continuously with the region’s integrated staleness. Evidence: Michon 2019 (graded replay counts scaling with reward magnitude); Mattar & Daw 2018 (EVB is a continuous function). This is the SYNTHESIS Q4 resolution above. Concretely: priority weights passed to the sleep-phase sampler are staleness[r] directly (or a monotonic transform), not staleness[r] > theta_sleep_priority ? 1 : 0.

3. Timing: FIRST-BOUT, stateless on replay-delay semantics. High-staleness regions should appear in first-bout sleep replay (i.e. the sampler consumes staleness priority immediately once the offline-mode gate permits it), not after a latency period. Downstream integration may still be delayed on the consumer side (Schlichting & Preston 2015 shows integration favours already-established memories, suggesting MECH-275 aggregation may have its own timing) — but that is a consumer-side property of MECH-273/MECH-275 schema-update, not of MECH-285’s sampling. MECH-285 does not need to track “how long since this region was last replayed” as an internal state.

4. Salience interaction: DISSOCIABLE ARMS converging at the sampler. Staleness priority (MECH-285) and dopaminergic salience priority (McNamara/Dupret 2014) are independent signals combined at the SWR-selection stage, not a single weighted score. Evidence: Joo & Frank 2018 multi-mode framing of SWR subpopulations; dissociable clinical phenotypes (PTSD salience-only vs schizophrenia staleness-only vs depressive rumination). Implementation: the sampler receives two priority vectors and combines them (default: additive with unit gains; tunable gains for lesion experiments). Do NOT pre-compose into a single weighted priority score.

5. Consumer scope: consolidation-SWRs only, not waking decision-support. MECH-285 writes its priority tag in a way the sleep-phase consumer (MECH-275 Bayesian aggregation, MECH-273 self-model aggregation) reads, but it does NOT bleed into waking retrieval-SWRs or awake decision-support channels. Those channels are priority-weighted by task demand (Mattar & Daw’s “need” term) and by DA salience, on different SWR subpopulations. The staleness map is gated to the sleep consumer — anchor key + consume_during_sleep_only=True semantics.

Substrate hook (deferred implementation):

  • ree_core/hippocampal/sleep_replay_sampler.py — new module (name TBD). Reads StalenessAccumulator.snapshot() at sleep-mode entry (gated by MECH-286 override-permitted transition); builds broad-unbounded seed pool across active + recently-inactive anchors; draws replay start-states with probability proportional to staleness[r] plus (additively) the DA-salience tag from the McNamara/Dupret arm; emits SWR-equivalent events consumed by MECH-273 / MECH-275 aggregators.
  • Configuration flag HippocampalConfig.use_mech285_sleep_priority (default False) so the offline arm can be activated independently of the online arms.

Falsification prediction restated with this design: with MECH-285 enabled and the DA-salience arm disabled, first-bout sleep replay of a low-arousal-high- novelty session should show start-state density concentrated on the novelty-region’s anchor keys with density proportional to their accumulated staleness — not a step change at some threshold, not time-delayed beyond the first bout, and not bleeding into subsequent waking decisions. Joint lesion of MECH-285 + MECH-094 should produce the rumination architecture (staleness accumulates from the traumatic segment but priority signal is broken and hypothesis-tag loss prevents whatever replay does occur from clearing the staleness tag).

Open design questions

  1. Trigger substrate canonical choice. LC vs DA vs LHb. The current architecture commits only to “broadcast event exists.” For substrate implementation, the simplest first pass treats trigger as a single computed signal combining the three; future refinement could split them. Lit-pull candidate (deferred): Chandler 2014 / Schwarz 2015 LC modular projections to test schema-targeting.

  2. attribution_weight parametrisation. The simplest first pass weights all regions in active_anchor_set equally. A more biological version weights by which streams contributed (z_harm vs z_goal). A more learned version uses an attribution head that trains during sleep on hindsight residuals. Open whether to specify in the substrate first pass.

  3. theta_reset(r) source. Per-region threshold or single global. Per-region is architecturally cleaner (different schema regions have different normal staleness levels) but requires per-region threshold initialisation. Default first pass: single global threshold.

  4. MECH-285 coupling magnitude (the SYNTHESIS Q4) — RESOLVED 2026-04-24 by targeted_review_mech285_sleep_replay_seed/ lit-pull (10 papers + SYNTHESIS.md). Quantitative/proportional coupling, not threshold-gated. Michon et al 2019 (Curr Biol) show replay count scales with reward magnitude as a graded signal, not a stepwise one above a criterion; Mattar & Daw 2018 (Nat Neurosci) provide the EVB normative backbone in which priority is a continuous expected-value-of- backup function. The staleness map therefore feeds replay priority as a scalar weighting (priority proportional to integrated waking staleness), not as a threshold gate. See “MECH-285 offline arm” design subsection below for the implementation commitments informed by the same lit-pull.

  5. Interaction with MECH-094 hypothesis tag. Anchor-reset events should probably be tagged with MECH-094’s hypothesis tag so that the new anchor’s first proposals are routed as probes (per MECH-269 probe channel) until the new anchor’s V_s accumulates sufficient anchor-eligibility evidence. Open whether this happens automatically or needs explicit specification.

  6. Failure mode to surface as diagnostic. If theta_trigger is set too low, every small surprise fires the trigger and the accumulator saturates immediately, producing anchor-reset chaos. Architecturally this matches the clinical observation of high- anxiety states with “everything seems wrong.” Worth surfacing as a diagnostic counter (mean trigger fire rate; mean anchor reset interval) rather than relying on the parameter sweep to catch it.

Status log

  • 2026-04-22 — Design doc written. MECH-287 reserved. Discussion origin: V3-EXQ-475 evidence (5-6 freeze releases, ~12 re-commits per release, 1000/1000 freeze active) shows that SD-036 decay + MECH-279 freeze exit are insufficient without an upstream anchor invalidation signal. Lit-pull on waking-phase online V_s invalidation (targeted_review_waking_v_s_invalidation/, 12 papers) converged on a two-circuit architecture: LC/DA broadcast trigger + OFC local accumulator. Recommendation is to register MECH-287 separately rather than expand MECH-284, because failure modes dissociate clinically and substrate readiness differs. MECH-269 refined with dual-trace anchor-reset criteria (Bouton 2004); MECH-272 promoted to load-bearing status for the fix because routing has to maintain anchor mixtures.
  • Registration in claims.yaml follows in same session.
  • 2026-04-22 (Phase 2 update) — targeted_review_v_s_foundation/ SYNTHESIS integrated. Five verdicts incorporated: (1) schema/event-segment as default V_s region granularity (NOT place-cell-field), with multi-scale support; (2) MECH-269 dual-trace must support both hard-switching (Wills 2005) and soft re-weighting (Colgin 2008) modes; (3) per-stream V_s reframed as projection-readout of integrated mixed-selectivity code (Eichenbaum 2017), validating the Phase 1 identity-proxy decision and ruling out per-stream forward predictors as a Phase 2 task; (4) anchor type catalogue supports both pure-place and stream-mixture-conditioned anchors (Wood 2000, Frank 2000) with ratio emergent from task; (5) MECH-287 trigger refined to dual-component coupled-stage architecture — upstream anchor-side CA1/CA3 mismatch comparator (Vinogradova 2001) routed via subiculum (O’Mara 2009) into the HC-VTA loop (Lisman & Grace 2005) coupled to the downstream broadcast (LC / DA / LHb). MECH-287 entry in claims.yaml updated with dual-component title, extended functional_restatement, and notes cross-reference to both lit-pulls. Region granularity also flagged as a tunable scaling lever per project memory project_intelligence_scaling_levers.md (Daniel’s 2026-04-22 aside on granularity / latent dim / network depth / sensory granularity coupling).
  • 2026-04-24 (Phase 3 online arm IMPLEMENTED) — MECH-284 staleness accumulator landed in ree-v3 as ree_core/hippocampal/staleness_accumulator.py, wired into HippocampalModule.tick_anchor_set between consume_boundary_events (Phase 2(ii)) and tick_hysteresis. Integration drains the broadcast event queue (MECH-287 events) with attribution_weight credit over the active anchor set; tick_leak applies EMA decay leak_factor=0.995 per call. Region key is (scale, segment_id) — matches per_region_vs partition; stream_mixture dropped because the accumulator indexes on the event-segment partition the Phase 2 per-region aggregation already uses. Two new HippocampalConfig flags, both default False: use_staleness_accumulator (instantiates the accumulator for diagnostic read-out without changing hysteresis behaviour) and use_mech284_hysteresis (additionally swaps MECH-269 AnchorSet.tick_hysteresis from the Phase 2 internal proxy (tick − last_accessed) * staleness_rate to the MECH-284 accumulator lookup). Both threaded through REEConfig.from_dims. Attribution modes: "equal" (baseline, default) and "stream_overlap" (cosine similarity between broadcast source_sources and anchor’s stream mixture). Dual-trace mark_inactive (NOT erase) preserved from Phase 2 — Phase 3 only changes how the hysteresis threshold is evaluated, not the extinction semantics. MECH-094 respected via call-site scoping: the accumulator is only touched during waking via HippocampalModule.tick_anchor_set, which is the realised-observation path; simulated/probe content goes through different write-gates and does not enter the broadcast queue. Backward-compat gate: 91/91 preflight + contract regression PASS with flags OFF (bit-identical behaviour vs Phase 2). Activation smoke (/tmp/smoke_mech284.py) covers three ARMs: ARM0 backward-compat (accumulator None when flag OFF); ARM1 integrate+leak (two-region equal-credit integration gives 0.4975 each after one broadcast and one leak tick, matching 0.5 × 0.995 exactly); ARM2 hysteresis swap (with staleness forced high, MECH-269 tick_hysteresis fires mark_inactive at k=3 under the MECH-284 lookup). V3-EXQ-478 queued as the Phase 3 diagnostic validation (2 seeds × 2 arms OFF vs ON; metrics: freeze_recommit_count, anchor_reset_count, mean_staleness_peak, action_class_ entropy). Phase 3 offline arm (MECH-285 sleep-priority readout) explicitly deferred — the infrastructure (per-region accumulator, snapshot(), get_stats()) is in place but the sleep-phase consumer is not yet wired.
  • 2026-04-24 (MECH-285 offline-arm design fold-in) — targeted_review_mech285_sleep_replay_seed/ lit-pull (10 papers + SYNTHESIS.md) commissioned to adjudicate the offline-arm design before implementation. MECH-285 literature_confidence 0 -> 0.866 across 11 entries. Five verdicts resolved: seed-coverage BROAD-UNBOUNDED (experience-structured), priority shape STALENESS-PROPORTIONAL (not threshold-gated — resolves Q4 above), timing FIRST-BOUT (integration delay is consumer-side not sampler-side), salience interaction DISSOCIABLE ARMS converging additively at the sampler, dual-trace at replay SUPPORTED. Q4 marked resolved in the Open Design Questions section. New “MECH-285 offline arm — design” subsection added specifying the 5 implementation commitments and the deferred substrate hook (ree_core/hippocampal/sleep_replay_sampler.py, flag HippocampalConfig.use_mech285_sleep_priority). Implementation still deferred — this commit is a design fold-in, not a substrate landing. V3 experiment proposal for the Q4 coupling test ( v3_exq_NNN_waking_load_to_sleep_priority.py) added to experiment_proposals.v1.json with gate requiring the offline-arm substrate to be wired first.

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