Hippocampal Anchor Selection via Regional Verisimilitude

Claim Type: mechanism_hypothesis (candidate, V3-pending) Scope: Anchor-state selection for hippocampal trajectory proposals issued between committed actions Depends On: ARC-007, ARC-018, MECH-089 (theta-gamma nesting), MECH-092 (quiescent SWR replay), MECH-094 (hypothesis tag), SD-005 (z_self/z_world split), SD-010/SD-011 (harm stream separation) Status: candidate Claim ID (reserved): MECH-269 Registration status: NOT YET in claims.yaml — deferred pending release of active implement-amygdala-analog claim on that file (2026-04-21) Implementation phase: V3

The gap this closes

ARC-018 states that the hippocampus constructs counterfactual rollouts from “the current latent state.” ARC-007 states that path memory indexes experienced trajectories through latent space. Neither document specifies which slice of the current latent actually becomes the rollout anchor. The proposer is described as starting “from now” — but “now” is not a single point in a distributed latent. It is a family of stream-local estimates (z_world, z_self, z_harm, z_goal, …) whose alignment with perception varies moment-to-moment.

Between committed actions — during the quiescent E3 heartbeat cycles that carry SWR-equivalent replay (MECH-092) — the proposer is expected to generate useful candidate trajectories. If it anchors on a low-verisimilitude region, the rollout carries a hidden lie forward: the starting point is fiction, so the trajectory is fiction about fiction. If it refuses to anchor at all on low-verisimilitude regions, the agent becomes incapable of probing the very territory where its model is weakest — exactly the territory that most needs exploration.

MECH-269 supplies the missing selection mechanism and splits proposals into two mechanistically distinct kinds: anchored and probed.

The mechanism

1. Regional verisimilitude as a per-stream quantity

Verisimilitude is the alignment between E1/E2 predicted latent and the actually realized latent, measured separately per stream. For each stream s ∈ {world, self, harm, goal, ...}, the proposer has access to a running alignment score:

V_s(t) = align(ẑ_s(t | t-Δ), z_s(t))

where ẑ_s(t | t-Δ) is the prediction issued Δ steps ago for stream s, and z_s(t) is the stream’s current realized value. Concretely this can be instantiated as a short-window EMA of the cosine similarity (or negative prediction error) per stream. MECH-098’s reafference cancellation feeds V_world specifically; the other streams use their own prediction/realization comparison.

2. Anchor eligibility

A stream is anchor-eligible at time t if V_s(t) ≥ θ_anchor(s). The proposer constructs its rollout starting-state by composing only the anchor-eligible slice of the current latent. Misaligned streams are held at their last-verified snapshot (an “anchor carryover”) rather than allowed to contaminate the rollout. This gives a principled per-stream construction of “now.”

3. Probe mode (the exploration escape valve)

A pure anchor-eligibility gate would suppress proposals in the very regions the agent needs to understand. The probe channel inverts the gate for a minority of proposals:

A rollout is issued as a probe when it is explicitly seeded from a low-verisimilitude region (or from a hybrid where one stream is a low-V probe seed and the others are anchored).
Probes carry a strengthened hypothesis tag (MECH-094) — they are marked as more fictional than anchored rollouts, not less. Their write-gate downstream is tighter; their role is information acquisition, not viability mapping.
Probe candidates come from curiosity / novelty signals (e.g. stream-local prediction-error spikes) rather than from verisimilitude. In effect: “I cannot currently model this region; deliberately simulate into it to find out what it contains.”

Anchored proposals update the viability map (ARC-018). Probe proposals do not update the viability map until a subsequent realized trajectory validates them; they update curiosity / epistemic-coverage structures instead.

4. Anchor/probe ratio under the heartbeat

The ratio between anchored and probe proposals is itself modulated by the control-plane heartbeat (ARC-023) and by global precision / mode. High-harm / high-commitment contexts push the mixture toward anchored-only (probe is cognitively expensive and can surface intrusive simulation — see MECH-094 failure modes). Low-threat / high-drive contexts widen the probe fraction. This falls out naturally from existing dACC / SalienceCoordinator gating and does not need a new controller.

Why regional, not global, verisimilitude

A global “overall model trust” scalar cannot do this job, for two reasons:

Dissociable streams. After SD-010/SD-011, z_world and z_harm_s can have very different alignment quality at the same moment (e.g. novel terrain is a z_world prediction failure but says nothing about z_harm tracking). Collapsing them wastes the decomposition the substrate is built to provide.
Partial anchoring is useful. In practice the proposer wants to roll forward from “current z_self + current z_harm + carried-forward z_world” when the world stream is temporarily unreliable but the agent’s own state and affective load are well-tracked. A global gate would force all-or-nothing anchoring.

This is the same logic that made per-claim evidence_direction necessary (see REE_assembly/CLAUDE.md): one verdict for a multi-part measurement is lossy.

Relationship to existing claims

Existing claim	Relationship
ARC-018 (hippocampal rollouts / viability mapping)	MECH-269 specifies how the starting state is chosen for the rollouts ARC-018 describes.
MECH-092 (SWR-equivalent replay in E3 quiescence)	Quiescent replay runs on anchored trajectories under MECH-269; probes do not enter the viability map via replay.
MECH-094 (hypothesis tag as categorical write gate)	Probe proposals carry a strengthened tag; tag loss on a probe is the candidate mechanism for confabulatory planning, consistent with existing confabulation framing.
MECH-098 (reafference cancellation)	Provides the per-step input to `V_world` specifically.
MECH-089 (theta-gamma nesting)	The per-cycle “snapshot” the proposer anchors on is the theta-cycle-averaged summary, not the instantaneous latent. Regional verisimilitude is computed on these summaries.
SD-005 (z_self / z_world split)	Precondition: MECH-269 is only meaningful once streams are separated.
SD-010 / SD-011 (harm stream separation)	Same — gives `z_harm_s` its own V.

Predicted observables (V3 scope)

A V3 experiment validating MECH-269 would measure:

Anchor-eligibility correctness: the proposer should refuse to anchor on streams whose subsequent realized alignment is below θ_anchor, and rollouts so-anchored should produce smaller prediction-vs-outcome gaps than rollouts that anchor indiscriminately.
Probe yield: probe-tagged rollouts should preferentially target high-PE regions and should not populate the viability map until validated by realized experience.
Streamwise dissociation: the agent should be able to plan under partial-anchor conditions (one stream unreliable) and should fail gracefully — not catastrophically — as compared to a no-anchor-selection baseline, which should confabulate forward from corrupted starts.
Heartbeat modulation of anchor/probe ratio: manipulating salience / harm load should shift the observed mixture, without re-training.

Candidate experiment name (not yet queued): v3_exq_NNN_mech269_anchor_probe_dissociation.py. Requires substrate hooks for per-stream V_s logging in HippocampalModule.propose_trajectories and an anchor/probe label field on emitted candidate trajectories.

Open design questions

Threshold setting. θ_anchor(s) per stream — hand-tuned, learned, or derived from the same precision signal that drives MECH-098 and dynamic-precision weighting (ARC-016)? First pass: tie it to a running quantile of the stream’s own prediction-error distribution, so it is intrinsically adaptive.
Carryover decay. How long can an anchored stream “carry over” its last-verified snapshot before it itself becomes untrustworthy? Probably governed by stream-local forgetting rates already implied by MECH-089’s theta-cycle horizon.
Probe quota. Is the probe fraction a hard budget per heartbeat window, or purely signal-driven? Budget is easier to implement and audit; signal-driven is more biologically plausible. Default to budget for V3 and revisit.
Interaction with post-commit vs pre-commit channels (the MECH-094 split). Probes are intrinsically pre-commit-like (rehearsal under tag); this may simplify rather than complicate the tag accounting.
Failure mode to watch for. If probe-tagged rollouts leak into the viability map because of tag erosion, the map becomes populated with fictional affordances. This is the architectural signature of confabulated planning and should be surfaced as a diagnostic counter, not just trusted to not happen.

Biological routing of anchored vs probe replay

The anchor/probe distinction is not purely functional. It has a candidate physical substrate in the known fan-out of ephaptically-coordinated hippocampal replay.

Coordination substrate. In CA1 and CA3, pyramidal-layer cell-body density is high enough that the local field generated by the firing population re-modulates that population’s own spiking. During sharp-wave ripples this ephaptic field effect is the dominant synchronising mechanism — tighter than synaptic transmission alone can produce (Anastassiou & Koch 2011; Jefferys; Buzsáki on SWRs). MECH-269’s per-stream verisimilitude readout is, in this framing, plausibly implemented as stream-local field coherence: streams whose predictions align with realized values produce coherent ensembles whose fields constructively reinforce; misaligned streams produce incoherent fields. “Anchor-eligible” reduces to “field-coherent at read time.”

Fan-out targets. Ephaptically-synchronised CA1/CA3 pyramidal output projects to a fixed set of downstream sites, and these map one-to-one onto REE modules currently in place or under construction:

Biological projection	REE module
Subiculum → entorhinal deep layers → neocortex	E1 world-model consolidation writes
Medial/lateral PFC (direct from ventral CA1)	SD-033a lateral-PFC analog (MECH-267 mode-conditioning consumer)
Basolateral amygdala (ventral CA1 → BLA)	BLA analog (under active implementation 2026-04-21)
Nucleus accumbens / ventral striatum	E3 commitment / viability evaluation
Mammillary bodies via fornix (Papez)	flagged for investigation (see `project_papez_circuit_investigation`); not yet a module

Routing hypothesis. The anchor/probe distinction predicts different downstream destinations, not a shared destination with a flag attached:

Anchored replay (high regional verisimilitude, trusted starting state) preferentially writes to subiculum → EC → neocortex and SD-033a lateral PFC. These are the destinations that update stored world structure and the viability map (ARC-018). Writing here commits the replayed content to long-term use.
Probe replay (low-V seed, strengthened MECH-094 hypothesis tag) preferentially engages BLA (affective tagging without consolidation) and NAc curiosity/novelty circuits. These destinations do not update the viability map until a subsequently realized trajectory validates the probe.

Under this framing, MECH-094 (the hypothesis tag) is not an undifferentiated “this is simulated” flag riding on top of otherwise identical replay. It is instantiated as a routing difference — probes reach a different subset of downstream consumers. Tag loss (confabulatory planning, PTSD-adjacent phenomenology) then has a specific biological signature: probe content reaching consolidation destinations it should not reach (EC/neocortex/lateral PFC). This is an auditable diagnostic, not just a hypothesis about internal state.

Why this matters for substrate implementation. The MECH-269 claim originally read as a purely computational gate inside HippocampalModule.propose_trajectories. The biological routing hypothesis says the gate should instead be realized as which downstream consumers receive the replay output. In V3 substrate terms: anchored rollouts publish to the channels consumed by E1 consolidation, SD-033a lateral PFC, and the viability-map writer; probe rollouts publish to the channels consumed by BLA (affective tagging) and the curiosity/novelty scaffolding. Consumers remain responsible for whether they read — the gate is at routing, not at a single source-side boolean.

Separate-claim candidates implied by this section. Two further MECH entries are implied and should be registered alongside MECH-269 rather than folded into it:

MECH-270 (candidate): ephaptic field coherence as the physical substrate for per-stream verisimilitude readout. Keeps the function/substrate distinction clean — same failure mode as SD-003 and SD-010/011 if we conflate them.
MECH-271 (candidate): hypothesis-tag realization as differential downstream routing rather than source-side flag. Anchored-to-consolidation vs probe-to-curiosity is the observable signature.

These are noted here but not drafted as separate docs this turn.

State-gating: sleep vs waking informational needs

The routing architecture of MECH-271 is silent on brain state. A conversation revision (2026-04-21) adds the commitment that the routing itself is state-gated, because sleep and waking serve distinct informational needs and therefore exploit the anchor/probe split differently.

Waking regime. The anchor channel dominates. The hippocampus is being asked to support decisions — give me rollouts from where I actually am to where I might go, and do it with high fidelity to the current world state. Probe events happen (DMN-style micro-quiescence, MECH-092) but they are minority events under a strong hypothesis tag (MECH-094) that prevents their content from writing to commitment-relevant consumers. The anchored / probe balance tips toward anchored.

Sleep regime. The probe channel has the floor. The computational work changes: it is no longer “make this decision” but “revise the schema that decisions run against.” That is full-Bayesian restructuring — testing which action-outcome attributions held up across episodes, creating new buckets for experiences that did not fit existing schemas, revising causal graphs where the data has accumulated evidence for or against prior attributions. Probe content is recruited heavily because novel combinations need to be tested; the hypothesis tag remains on to preserve the distinction between tested-and-accepted (consolidation-routed) and tested-but-not-committed (remains in repertoire).

Tang et al. 2017 offer the empirical hint: awake and sleep SWRs differ in their cortical reactivation quality. Our architectural reading is that the difference is not incidental — it reflects the functional regime switch between decision-support (awake, anchor-dominant) and schema-revision (sleep, probe-dominant Bayesian work).

Registered as MECH-272. State-gated routing. Waking = anchor-dominant fan-out bias; sleep = probe-dominant fan-out with Bayesian restructuring of the schema repertoire. Falsifiable (primary): V3 experiments that compare replay-to-consumer routing during wake vs simulated-sleep phases should show the dominance shift. Falsifiable (secondary): under state-gating loss (e.g., REM-like protocols run in waking state), routing should lose the regime-appropriate bias.

Sleep-dependent self-model aggregation

The self-model has a waking half and a sleep half. The waking half is SD-003: E2’s counterfactual causal signature E2(z_t, a_actual) − E2(z_t, a_cf), attributing the discrepancy between predicted and actual outcome to the agent’s own action. This is probabilistic and single-episode. It gives us “in this episode, this action looks self-initiated with probability p.”

The sleep half is the aggregation that turns many single-episode attributions into a stable self-model. Single-episode attribution misses:

Delayed consequences where an action-outcome link spans more episodes than a single waking rollout can encompass.
Failed counterfactuals where the counterfactual action would have produced the same outcome — the action felt voluntary but was not actually causally efficacious.
Systematic attribution biases that a single episode cannot distinguish from noise but aggregate evidence across episodes can detect and correct.

Full-Bayesian aggregation of SD-003 outputs during sleep, routed by the anchored channel to E1 consolidation (world model revision) and to SD-033a (viability-map revision of which actions the agent can credit to itself), is how REE builds a stable sense of self. Without it, SD-003 gives the agent only an episode-local causal signature — not a durable self.

Registered as MECH-273. Sleep-dependent aggregation of single-episode SD-003 self- attributions into a stable self-model. Falsifiable (primary): ablating the sleep-phase aggregation step should leave single-episode self-attribution intact but degrade stability of self-attribution across episodes — concretely, the agent should be less able to correct previously-held spurious self-attributions when post-hoc evidence accumulates against them. Falsifiable (secondary): the pattern should generalise — any attribution whose stable form requires multi-episode Bayesian revision should show the same sleep-dependence.

INV-049 (the mathematical necessity of offline phases for model-building agents) sits directly under this claim. This is the specific content offline phases are doing for the self-model case.

Other-attribution sleep dependence (V4-reserved)

The same architectural pattern extends to other-attribution when V4 adds it. V4’s fast empathy system (ARC-010, MECH-217 if wired) produces single-episode attributions of other-agent action-outcome links: “agent j did action a, which looks causally responsible for outcome o.” By the same argument as MECH-273, these single-episode attributions need sleep-phase Bayesian aggregation before they stabilise into a model of agent j’s dispositions and causal powers. Without it, V4’s other-model is episode-local — which is not what a stable theory of other minds requires.

Registered as MECH-274, flagged V4. The extension is architecturally parallel to MECH-273 but operates on other-agent attributions rather than self-attributions. Not for V3 implementation; flagged here to reserve the mechanism and prevent V4 from rediscovering it independently.

Anchor reset criteria (V3-EXQ-471 motivated extension, 2026-04-22)

The original MECH-269 specification gave per-stream verisimilitude a role in initial anchor selection: which slice of the current latent is eligible to seed a rollout at the moment the proposer fires. It did not commit to a reset criterion — under what conditions an anchor that was valid at one heartbeat becomes invalid and must be discarded, even if no new commit has occurred to obviously reset state.

V3-EXQ-471 forced this question. The trace (full discussion in psychiatric_failure_modes.md “Catatonia, Subtype II”) shows the proposer locked onto an anchor formed at t=0 (the moment of a single hazard contact) and continuing to issue rollouts from that anchor for 200 steps, despite 190 steps of subsequent quiet during which the world had clearly moved on. Mode stayed avoid, action stayed locked, position stayed at [2,1], energy went to zero — and the proposer never reconsidered the anchor.

This is a real failure mode that the per-stream verisimilitude gate does not catch on its own. V_world was probably perfectly fine at t=190 (predictions matched realizations — both predicted and realized “still nothing happening”). The problem was not that the anchor’s streams had become unreliable; it was that the anchor was temporally stale — it had been formed under conditions (z_harm_a = 0.82, hazard adjacent) that no longer held by t=50, but nothing in the proposer noticed.

What the reset criterion must commit to

Anchor reset is required when the regulatory state of the streams the anchor was composed from has materially changed, even when those streams’ verisimilitude remains intact. Concretely:

Decay-driven reset. If a stream that was high-magnitude at anchor formation has decayed (per SD-036 GABAergic decay regulator) below a freshness threshold by anchor read time, the anchor for that stream is no longer trustworthy as a representation of the current mode-relevant state. The anchor was “I am in a high-harm context, plan accordingly”; the current state is “I am in a low-harm context”; the rollouts issued under the stale anchor are stale.
Mode-context reset. Anchors are formed under a particular operating mode (SD-032a current_mode). When the SalienceCoordinator’s mode_switch_trigger fires, anchors composed under the previous mode should be evicted from the eligibility pool — not merely augmented with new ones. EXQ-471’s pathology is precisely that no mode switch ever fired, so this criterion is necessary but not sufficient (the agent also needs the prior criterion).
Time-since-formation cap. A purely temporal cap (anchors older than N heartbeat cycles must be re-validated against current realization) is a defensive backstop against staleness that the regulatory-state criterion misses. Initial value: re-validate after ~10 theta cycles (~MECH-089 timescale) of being held without commit-driven update.

Interaction with SD-036 GABAergic decay

The anchor reset criterion is the read-side mirror of SD-036’s write-side decay. SD-036 decays the streams; the reset criterion ensures the proposer notices that the decay has happened. Without the reset criterion, SD-036 alone is insufficient: the streams could decay back to baseline while the proposer continued to issue rollouts from an anchor formed when they were high.

Architecturally this means:

SD-036 (regulator layer):  decays z_harm, z_beta toward baseline
        |
        v
MECH-269 reset criterion:  detects stream decay relative to anchor formation state
        |
        v
Anchor pool eviction:      stale anchors removed; proposer re-anchors on current state
        |
        v
Mode arbitration:           with current-state anchor, mode flip becomes possible

The chain is what unlocks the EXQ-471 pathology. SD-036 alone unlocks the streams but not the proposer; reset criterion alone has nothing to react to without decay; mode arbitration alone has nothing to switch on without a current-state anchor to evaluate. All three are required and they are mechanistically coupled.

What this implies for the design

The original design treated V_s(t) ≥ θ_anchor(s) as the only gate on anchor eligibility. The reset criterion adds a second gate:

anchor_eligible(s, t) =
    (V_s(t) >= theta_anchor(s))                           # original verisimilitude gate
    AND
    (regulatory_state_drift(s, t, t_formation) < theta_drift)   # NEW reset gate
    AND
    (mode_at_formation == current_mode)                   # NEW mode-context gate
    AND
    (t - t_formation < max_anchor_age)                    # NEW temporal backstop

regulatory_state_drift(s, t, t_formation) is a per-stream measure of how much the stream’s value has changed between anchor formation and current read, normalized by its plausible decay over that interval (so that “expected decay” is not flagged as drift, only “more decay than the regulator would have produced” or “decay despite formation conditions persisting”). The exact form is an open design question; first pass is |z_s(t) - z_s(t_formation)| / expected_drift(t - t_formation, tau_s).

Predicted observable

A V3 experiment validating the anchor reset criterion would re-run V3-EXQ-471 with SD-036 enabled and the reset criterion enabled. Predicted trace at seed 0 ep 0:

t=0: hazard contact, z_harm_norm jumps to ~0.82, anchor formed at this state
t=10–30: harm input ceases, z_harm_norm decays under SD-036
t=~30: z_harm_norm has decayed below freshness threshold relative to anchor formation; reset criterion fires; anchor evicted
t=~30+: proposer re-anchors on current low-harm state; rollouts now consider goal-seeking trajectories
t=~40–50: mode flip from avoid to a goal-seeking mode; agent resumes navigation
t=~50+: episode proceeds normally

Without the reset criterion, SD-036 alone is predicted to leave the agent in a strange intermediate state — unlocked streams but stale anchors — possibly producing intermittent or delayed recovery rather than the clean ~t=50 flip.

Open design question

What happens during sleep-mode replay (MECH-272)? Sleep-mode probe-dominant routing deliberately uses anchors that may be temporally stale or counterfactual. The reset criterion as specified would suppress sleep-mode probe replay, which is wrong. The likely resolution: the reset criterion applies to anchored proposals only; probe proposals are explicitly exempt because their job is to seed from non-current states. This preserves sleep-mode Bayesian restructuring (MECH-272) while preventing waking-mode anchor staleness. This may need its own MECH ID if the anchor/probe distinction in reset behaviour proves architecturally substantive.

V_s as bidirectional signal: runtime gate + sleep priority (MECH-283 / 284 / 285)

The V_s signal introduced for MECH-269 has so far been used in one direction only: as a runtime gate on anchor proposal eligibility. But the same per-stream verisimilitude scalar carries information that the architecture is currently leaving on the floor. This section registers three further uses, each as its own MECH:

MECH-283: V_s recognition-for-recall gate

Statement: V_s computed at retrieval time (not proposal time) gates whether a stored episode is even a candidate for re-entry into the proposal stream. Recognition (“does this fit my current situation?”) is logically prior to anchor-eligibility (“is this a good place to start a rollout from?”).

Distinction from MECH-269: MECH-269’s V_s is computed against the proposed anchor state; MECH-283’s V_s is computed against the retrieved episode content. A retrieval candidate with low V_s on its episode content fails recognition before its anchor is even tested. This explains why hippocampal recall is content-coherent even when the anchor would technically be eligible — the recognition gate fires first.

Failure mode: Loss of MECH-283 produces intrusive recall — episodes get retrieved that don’t fit current context (PTSD flashback phenomenology, schizophrenic loose association). Note that this is a different failure surface from MECH-094 hypothesis-tag loss (which produces simulation→reality confusion), though both can co-occur.

Substrate prediction: A retrieval-time V_s computation against episode content, upstream of the anchor-eligibility gate of MECH-269. Architecturally this is a small addition to the hippocampal proposer’s retrieval pipeline.

MECH-284: V_s residual schema-staleness accumulator

Statement: When an anchored rollout is committed and executed, the discrepancy between the rollout’s predicted downstream V_s trajectory and the realised V_s trajectory accumulates as a per-region staleness signal on the schema regions the rollout traversed. Persistent low V_s on a schema region — i.e. the schema repeatedly under-predicts current state — tags that region as needing revision.

Distinction from generic prediction error: Standard prediction error (E1 sensorium loss, E2 transition loss, E3 harm/goal loss) operates on the latent values themselves and updates the relevant heads via gradient. MECH-284 operates on the schema regions in the hippocampal map, accumulating across episodes, and is read by sleep consolidation rather than by online learning. It is a slow, region-indexed “this part of my map is wrong” signal that can persist across many waking episodes without prompting immediate parameter updates.

Failure mode: Saturated staleness map produces global “everything is uncertain” phenomenology — anxiety-disorder-analog over-attention to schema revision needs. Absent staleness map produces the opposite — confident misperception, schema rigidity under environmental change (some autism-spectrum / OCD-spectrum phenomenologies map here).

Substrate prediction: A region-indexed accumulator (one float per schema region) that integrates (predicted_V_s - realised_V_s) across executed rollouts, with a slow leak so very old residuals decay. Read-only during waking; consumed during sleep by MECH-285.

MECH-285: Sleep-consolidation priority from V_s residuals

Statement: During sleep-phase Bayesian aggregation (MECH-275) and sleep-dependent self-model aggregation (MECH-273), the order and weight of replay events is biased by the schema-staleness map produced by MECH-284. Replay is not uniform across the schema and is not solely salience-driven (dopaminergic tagging); it is also prioritised by accumulated verisimilitude residuals. The schema regions that have accumulated the most “this is wrong” evidence get replayed first, more often, and with higher schema-update authority during sleep.

Distinction from existing replay-prioritisation accounts: The dopaminergic salience-tagging account predicts that emotionally significant events get prioritised. MECH-285 predicts an additional, dissociable priority signal: epistemically important regions (where the model is wrong) get prioritised even if they are emotionally flat. The two priority signals are partially redundant in normal life (because emotionally significant events are often also epistemically important) but dissociable under specific conditions: a calm, low-arousal but high-novelty environment should produce sleep-replay biased toward novelty even without salience tagging.

Failure mode: Loss of MECH-285 produces PTSD-like rumination — the staleness tag accumulates from a traumatic episode but never gets cleared by sleep replay, because the priority signal that should pull that region into replay is broken. Hypothesis-tagged simulation (MECH-094) prevents the replay that does occur from updating the schema, locking the rumination loop. This is the architecturally clean PTSD failure mode — a co-occurrence of MECH-285 priority loss and MECH-094 hypothesis-tag dysfunction.

Substrate prediction: A read of the MECH-284 accumulator at sleep onset that biases the SWR-equivalent replay scheduler. Concretely: sleep replay’s start-state distribution (from MECH-269 / Pfeiffer & Foster 2013 substrate) is weighted by accumulated staleness. Updates the existing depends_on of MECH-272 (state-gated routing), MECH-273 (sleep self-model aggregation), MECH-275 (sleep-phase Bayesian aggregation general mechanism) to reference MECH-285.

Bidirectional summary

V_s carries information in two directions through the architecture:

Direction	Mechanism	Phase	Use
Forward (gate)	MECH-269	Waking	Anchor-eligibility for proposal stream
Forward (gate)	MECH-283	Waking	Recognition-for-recall (retrieval candidacy)
Reverse (residual)	MECH-284	Waking	Per-region schema-staleness accumulator
Reverse (priority)	MECH-285	Sleep	Replay scheduling and schema-update authority

The architectural unification: V_s is the schema’s self-monitoring signal. During waking it gates what the schema is allowed to propose and recall; during waking it also accumulates evidence about its own under-fit; during sleep it directs the repair. This is the symmetry the user identified — verisimilitude is more than grounding the now; it is the schema’s principal mechanism for knowing what it does not yet know well.

Status log

2026-04-21 — Design doc written. Claim ID MECH-269 reserved but not yet in claims.yaml; registration deferred pending release of concurrent active claim on the registry file. Discussion origin: user-initiated exploratory question about anchor selection for inter-commit hippocampal proposals, linking verisimilitude to anchor eligibility and proposing the anchor/probe split as the answer to the exploration-suppression objection.
2026-04-21 (same session) — Biological routing section appended. Ephaptic coupling flagged as candidate substrate for regional verisimilitude readout; anchored-vs-probe distinction recast as a routing hypothesis into known hippocampal fan-out targets, mapped onto existing/active REE modules (E1 consolidation, SD-033a lateral PFC, BLA analog, NAc, Papez). MECH-270 and MECH-271 noted as implied separate claims.
2026-04-21 (later) — MECH-269/270/271 registered in claims.yaml. Targeted lit-pulls landed (Jadhav 2016, Girardeau 2017, Anastassiou 2011, Pfeiffer & Foster 2013, Dragoi & Tonegawa 2011/2013, English et al 2014, Tang et al 2017, Ólafsdóttir 2018, Buzsáki 2015, Foster 2017). MECH-269 literature_confidence 0.852; MECH-271 0.795; MECH-270 0.750.
2026-04-21 (later still) — State-gating / sleep-waking / self-model aggregation / V4 other-attribution sections added. MECH-272, MECH-273, MECH-274 registered. Discussion origin: user observation that sleep and waking serve distinct informational needs (sleep = Bayesian schema revision, waking = decision-support using existing schemas) and that the self-model has a waking half (SD-003) and a sleep half (aggregation of single-episode attributions).
2026-04-22 — Anchor reset criteria section added, motivated by V3-EXQ-471 catatonic lock trace (200-step avoid-mode lock with anchor stuck at t=0 harm event despite 190 steps of subsequent quiet). Reset criterion specified as the read-side mirror of SD-036’s write-side decay; introduces three additional anchor-eligibility gates (regulatory drift, mode-context, temporal cap). Identified as architecturally coupled with SD-036 (registered same session) — neither alone resolves the EXQ-471 pathology. Sleep-mode probe exemption flagged as open design question, possible candidate for new MECH if substantive.
2026-04-22 (later) — V_s bidirectional section added (MECH-283 / 284 / 285). Discussion origin: user observation that the verisimilitude function is doing more than grounding the now — it also carries information about recognition-for-recall, about where hippocampal plan predictions don’t match well, and about which schema-level representations need fine-tuning during sleep. Architectural payoff is the unification: V_s is the schema’s self-monitoring signal across both phases, with forward use during waking (gating proposals and retrieval) and reverse use accumulated during waking and consumed during sleep (schema-staleness map → replay priority → schema repair). PTSD failure mode now has a clean architectural account as MECH-285 priority loss co-occurring with MECH-094 hypothesis-tag dysfunction.