MECH-285 Sleep-Replay Seed-Distribution Lit-Pull: Narrow vs Broad Staleness-Priority Coverage

Created: 2026-04-24 Origin: MECH-284 Phase 3 (online arm) landed in ree-v3 on 2026-04-24 (ree_core/hippocampal/staleness_accumulator.py). MECH-285 (offline arm – sleep replay prioritisation weighted by accumulated staleness) is the remaining half of the V_s bidirectional cluster. Before implementing it one architectural decision needs biology validation: whether the sleep- replay start-state distribution should be biased only by active anchor staleness (narrow) or should also include recently-invalidated / Bouton- inactive dual-trace anchors (broad). The MECH-284 region map and the anchor-set dual-trace preservation are both already present; what needs settling is whether sleep replay reaches the inactive traces at all, and if so with what timing and priority. Prior V_s pulls – targeted_review_waking_v_s_invalidation/, targeted_review_v_s_foundation/, targeted_review_connectome_mech_269/, targeted_review_connectome_mech_270/, targeted_review_connectome_mech_271/ – covered waking invalidation, schema-region granularity, multi-map coexistence, and routing. Gap: nothing yet directly covers the start-state distribution of sleep-phase replay after remap / extinction, which is the MECH-285 implementation pivot.

Prompt

/lit-pull Hippocampal sleep-phase replay start-state distribution after anchor invalidation / context change. Three converging architectural questions for MECH-285 (V_s residual-staleness sleep-priority readout): (1) seed coverage – during sleep replay, does the start-state distribution draw only from currently-active (winning) place-cell ensembles, or does it also draw from recently-extinguished / remapped / dual-trace-inactive contexts? (2) staleness-proportional vs threshold-gated priority – does replay priority scale quantitatively with accumulated epistemic staleness across regions, or is there a threshold above which a region enters the replay pool with roughly uniform priority thereafter? (3) timing of inactive-trace replay – if inactive traces are replayed at all, does this happen on the first sleep bout after invalidation, or with a delay (e.g., first-night consolidation of the winning trace; subsequent-night integration of the loser)?

The architectural commitment to test: MECH-285 biases sleep-replay start-state sampling by accumulated MECH-284 staleness over active anchors only (narrow reading), with staleness-proportional priority and no delay. The alternative is that inactive anchors retain replay eligibility for some window post-invalidation (broad reading), in which case MECH-285 must maintain a persistent region->seed map that survives AnchorSet.mark_inactive, and MECH-284 staleness must continue accumulating on inactive regions (or be re-routed to a separate inactive-trace staleness channel) so the sleep consumer can rank them.

Specific neurobiological systems to cover

1. Sleep-phase replay start-state distribution – direct evidence
   • Pfeiffer & Foster 2013 (Science) – sequence replay start-state in goal-directed navigation; the canonical substrate Foster identified for biased sampling. (Already cited in targeted_review_connectome_mech_269/; re-use, do NOT re-pull.)
   • Olafsdottir, Barry, Saleem, Hassabis & Spiers 2015 (eLife) – replay of non-local and never-visited trajectories; establishes that the sleep-replay distribution is NOT limited to most-recent waking experience.
   • Olafsdottir, Bush & Barry 2018 (Current Biology) review of replay content.
   • Gupta, van der Meer, Touretzky & Redish 2010 (Neuron) – replay of untaken paths; priority signal extends to counterfactual trajectories.
   • Wu & Foster 2014 (J Neurosci) – reverse-replay magnitude scales with reward; establishes a non-trivial priority-by- relevance pattern in the canonical sleep-replay substrate.
2. Replay of remapped / recently-extinguished contexts
   • Liu, Gillespie, Swanson, Berkowitz & Frank 2021 (Nature) – hippocampal replay following aversive experience and context remapping; direct test of whether recently-invalidated place representations still appear in subsequent replay.
   • Karlsson & Frank 2009 (Nature Neurosci) – awake replay of remote experiences; establishes the time-window over which past contexts remain replay-eligible.
   • Pfeiffer 2020 (Hippocampus) review on content of replay.
   • Genzel, Dragoi, Frank, Ganguli, Redish, Tononi (2020) – review of replay-content literature including extinction contexts.
3. Timing of trace integration during sleep
   • Tambini & Davachi 2013, 2019 – cross-state replay persistence and forward propagation (already cited in SD-032a; cite where relevant, do NOT re-pull).
   • Schlichting & Preston 2015, 2017 – integration of previously- distinct contexts during sleep; whether newly-invalidated traces are integrated on the first post-invalidation sleep or only on subsequent bouts.
   • Lewis & Durrant 2011 (Trends Cogn Sci) – overlapping replay hypothesis; schema-integration timecourse across multiple sleep bouts.
   • Diba & Buzsaki 2007 (Nature Neurosci) – forward and reverse replay in sleep; whether the two modes sample start-states differently.
4. Priority signals in sleep replay – quantitative vs threshold
   • Michon, Sun, Kim, Ciliberti & Kloosterman 2019 (Curr Biol) – post-learning replay prioritisation by behavioural relevance.
   • Joo & Frank 2018 review (Nat Rev Neurosci) – hippocampal SWR content and priority signals.
   • Mattar & Daw 2018 (Nat Neurosci) – normative model of replay prioritisation (EVB); predicts staleness-proportional priority under a well-specified utility function.
   • Whether empirical evidence matches the quantitative prediction or supports a threshold-gated scheme.
5. Dual-trace extinction at the neural level
   • Radulovic, Jovasevic & Meyer 2017 – hippocampal context reinstatement after extinction; whether the extinguished trace remains accessible to replay.
   • Quirk & Mueller 2008 – extinction is new learning, not erasure; biological grounding for dual-trace preservation.
   • Bouton 2004 (already cited in MECH-269/284/285 registration; re-use, do NOT re-pull).
   • Whether the Bouton dual-trace at the behavioural level is reflected in sleep-replay patterns – does the inactive trace get replay bandwidth, and does that bandwidth track accumulated staleness?
6. Neuromodulatory gating of replay priority
   • Wang & Morris 2010; McNamara, Tejero-Cantero, Trouche, Campo- Urriza & Dupret 2014 – dopaminergic tagging of replay-eligible experiences; whether staleness competes with or composes with dopamine-tag priority.
   • Swift, Gross, Frazer, Bauer, Clark, Vazey, Aston-Jones, Li, Pickering, Sara & Totah 2018 (Current Biology) – LC-NE in sleep / post-learning consolidation.
   • Whether MECH-285 (epistemic-staleness priority) is dissociable from the salience-priority arm or is a composite.

Architectural questions the lit-pull should help answer

1. Seed-coverage (narrow vs broad). Is the sleep-replay start- state distribution confined to currently-active ensembles (narrow), or does it span recently-invalidated traces for some window (broad)? Narrow -> MECH-285 implementation is the active- anchor-staleness biased sampler we already have infrastructure for. Broad -> MECH-285 must persist a region->seed map across AnchorSet.mark_inactive, and the staleness accumulator must continue writing to inactive regions (or be mirrored on a separate inactive channel).

2. Priority shape. Is empirical priority staleness-proportional (quantitative) or threshold-gated (qualitative)? Both are implementable; the biology should fix the default. Quantitative implies a softmax / power-weighted sampler; threshold implies a categorical “stale-enough-to-replay” flag.

3. Timing. Does replay of recently-invalidated traces appear on the first post-invalidation sleep bout, or with a delay? First-bout -> MECH-285 can be stateless (just reads the current staleness snapshot). Delay -> MECH-285 needs a temporal-window parameter (minutes-to-hours biological; tick-count in-silico) governing when inactive traces become replay-eligible.

4. Interaction with salience. Dopamine-tag priority (salience) and MECH-285 epistemic-staleness priority are ostensibly dissociable (MECH-285 spec explicitly says so). Does the literature support two distinct priority signals, or do they compose into a single weighted score? If compositional, the MECH-285 sampler must take both as inputs and arbitrate; if dissociable, they are independent write paths on the same replay-selection stage.

5. Dual-trace at the replay-content level. MECH-269’s dual-trace preservation (Bouton) was validated at the anchor-set level by the V_s foundation pull. Does it also hold at the sleep-replay level – does an inactive trace still generate SWR content? If no, the architectural commitment that “inactive anchors are retrievable” is narrower than assumed and MECH-285 narrow-mode is the only consistent implementation. If yes, broad-mode is the faithful reading.

Output structure

Standard targeted_review_*/ format. Suggested directory: evidence/literature/targeted_review_mech285_sleep_replay_seed/

Per-paper records as usual (record.json + summary.md). After the pull, write a short SYNTHESIS.md flagging:

• Seed-coverage verdict: narrow (active-only) vs broad (active + recently-inactive), with timing window for the broad case.
• Priority-shape verdict: staleness-proportional vs threshold-gated, with default parameter ranges for the in-silico sampler.
• Timing verdict: first-bout vs delayed, with the delay distribution if applicable.
• Salience-interaction verdict: dissociable independent signals vs composed weighted score.
• Dual-trace-at-replay verdict: does SWR content include inactive traces?
• Implementation recommendation: narrow / broad-window-N / broad- unbounded, with one-line rationale tying the recommendation to the specific verdicts above.

Estimated scope: ~10 papers, single session.

Notes for the agent doing the pull

• The user is a consultant psychiatrist; clinical mappings welcome (e.g., PTSD intrusive-replay pattern as broad-coverage inactive- trace replay with MECH-094 tag loss; depressive rumination as narrow-coverage high-salience loop; novelty-blindness in schizophrenia as comparator-hypoactivity upstream of MECH-284).
• Re-use papers already pulled in prior V_s / connectome / routing / homeostatic-override pulls; cite where relevant, do NOT re-pull.
• The exemplar that motivates this pull is the MECH-285 implementation decision (narrow vs broad seed coverage) documented in docs/architecture/v_s_invalidation_runtime.md Status log entry 2026-04-24 (Phase 3 online arm IMPLEMENTED) and Open design questions section Q4 (MECH-285 coupling magnitude).
• Connect to existing claim cluster: MECH-285 (sleep-priority readout target), MECH-284 (upstream staleness substrate, Phase 3 online arm landed), MECH-269 (dual-trace anchor-set preservation), MECH-273 (sleep self-model aggregation, downstream consumer of biased replay), MECH-275 (general sleep Bayesian aggregation, downstream consumer), MECH-094 (hypothesis-tag gating of replay-write paths; co-failure with MECH-285 produces PTSD architecture per claim 20249-20245).
• Be alert to evidence that the architecture should be revised. If empirical timing shows delayed integration (second-night or later for inactive traces), MECH-285 needs a temporal-gating parameter not currently in the substrate plan. If priority is threshold- gated rather than proportional, the sampler topology changes (categorical flag vs weighted distribution). If narrow-coverage is supported, the V_s invalidation runtime design doc open- question Q4 can be closed with minimal further architecture work; broad-coverage substantially expands MECH-285’s state footprint.