D_V Temporal-Depth Verisimilitude — V3 Synaptic Form (ARC-054 V3 promotion)

Claim cluster: ARC-054 (D_V trajectory selection – promoted v4 -> v3 in rollout-horizon synaptic-EMA form), INV-068 (temporal depth requirement), MECH-269 (per-region V_s), MECH-284 (staleness accumulator). Registered: 2026-04-26 Status: candidate, implementation_phase: v3, v3_pending: true (V3 form); ARC-054 V4 form (with full ARC-053 TCL substrate + MECH-225 oscillatory multiplexing + MECH-228 ephaptic field-coherence) remains v4-deferred. Origin: thought_intake_2026-04-09_verisimilitude_ethics.md Thread A. Architectural commitment: the V3 working-model needs a temporal-depth signal that hippocampal rollout evaluation can integrate over its planning horizon. The V4 form requires phase-coherent substrate (TCL); the V3 form approximates this with a synaptic EMA over a content-similarity proxy. The V3 form is sufficient to close the temporal-depth gap that surfaced in EXQ-074b/076b path collapse and the EXQ-471/478/480/483/483a freeze- recommit oscillations. The V4 form refines, not replaces, the V3 form.

Why promote now (the gap)

ARC-054 in its full V4 form (J(pi) = sum_k gamma^k * V_hat_pi(t+k) over a phase-coherent TCL) is correctly v4-deferred – the substrate it depends on (ARC-053 distributed phase loop, MECH-225 oscillatory multiplexing, MECH-228 ephaptic field-level coherence) is not in V3 scope.

But MECH-269 V_s as currently implemented is instantaneous. The waking-arm circuit (MECH-287 broadcast trigger + MECH-284 staleness accumulator + MECH-269 anchor-reset hysteresis) tells E3 that an anchor’s verisimilitude has degraded; it does not tell E3 how the predicted future verisimilitude trajectory of a candidate rollout compares to the predicted future verisimilitude of an alternative. Several V3 failure modes implicate this missing horizon-integration signal:

  • EXQ-471/478/480/483/483a freeze-recommit cycles: SD-036 GABA decay releases the freeze gate, but E3 has no signal that the candidate trajectory drawn from the original avoid-anchor will predictably re-stoke z_harm_a within ~80 steps. Instantaneous V_s doesn’t carry the projection.
  • EXQ-074b/076b path collapse (MECH-124 z_goal salience diagnostic): z_goal salience is not competitive with harm salience because the rollout cost is instantaneous; a trajectory with high anticipated future D_V (sustained coupling to a goal-relevant schema) is not preferred over a trajectory with high instantaneous V but predicted collapse.
  • SD-029 monostrategy lock-in (V_s_baseline-monostrategy diagnostic EXQ-482): rollout evaluation cannot compare alternative operating modes when each mode’s verisimilitude is read at a single timestep.

The V3 D_V form closes this gap synaptically.

V3 form (rollout-horizon synaptic-EMA D_V)

Variables

For each candidate rollout pi proposed by the hippocampal proposer (MECH-269 anchored draw + MECH-285 sleep-priority probe draw, both already implemented), and for each step k = 0, 1, ..., H of the rollout (H = current planning horizon):

  • V_hat_pi(t+k) – predicted regional verisimilitude at step k of rollout pi.
  • D_V_pi(t+k) – predicted temporal-depth verisimilitude at step k.

Computation (V3 form, no phase substrate)

V_hat_pi(t+k) is the per-region V_s readout (already implemented as MECH-269 Phase 2 iii T4 per_region_vs[(scale, segment_id)][stream]) read at the predicted future region (scale, segment_id) that pi visits at step k. Source: the BoundaryEvent segmenter (MECH-288) projects predicted segment_ids forward over the rollout. For steps where the projected region is not in the AnchorSet, V_hat_pi(t+k) falls back to the per-stream V_s EMA (MECH-269 Phase 1) at the upstream anchor’s stream mixture.

D_V_pi(t+k) is an EMA over the rollout step:

D_V_pi(t+0) = V_hat_pi(t+0)
D_V_pi(t+k) = lambda * D_V_pi(t+k-1) + (1 - lambda) * V_hat_pi(t+k)   for k >= 1

with lambda configurable (arc054_dv_lambda, default 0.85 to match the typical hippocampal rollout horizon of ~10-30 steps).

Trajectory cost

J(pi) = sum_{k=0}^{H} gamma^k * D_V_pi(t+k) is added as an additive bias to the existing E3 trajectory selection score (which already incorporates MECH-014 harm-cost, MECH-016 goal-attainment, etc.). J(pi) is exposed as arc054_dv_score in the rollout metadata so governance experiments can measure the contribution.

Wiring

  • E3Selector reads per_region_vs via the existing HippocampalModule API.
  • HippocampalModule grows a small helper predict_rollout_vs(rollout) -> List[float] that walks the rollout and reads V_hat_pi(t+k) per step using the MECH-288 segmenter projections.
  • Master flag use_arc054_dv (default False, bit-identical-when-OFF preserved per the V3 substrate-flag convention).
  • Sub-knobs: arc054_dv_lambda (EMA decay), arc054_dv_gamma (rollout discount, default matches e3_planning_gamma), arc054_dv_score_weight (scalar coefficient on additive bias, default 1.0), arc054_dv_horizon_cap (truncate at min(H, cap), default 30).

What V3 deliberately does NOT include

  • No phase-channel projection (MECH-225 oscillatory multiplexing). V3 reads V_s as a scalar per (region, stream); V4 will read phase-aligned V(t) components.
  • No TCL phase-coherence substrate (ARC-053). V3 uses synaptic readout; V4 reads cross-frequency-coupled multiplexed signals.
  • No ephaptic field-level coherence support (MECH-228). V3 treats V_s changes as discrete EMA updates; V4 will treat them as continuous field-level coupling magnitude.
  • No D_self computation as a separate variable (INV-069 self-as-coherence-trajectory). V3 D_V is computed over the per-stream V_s readout, of which z_self is one stream; D_self is implicitly D_V restricted to stream = z_self. V4 may promote D_self to its own first-class signal.
  • No multi-agent extension (β_j coefficients, ARC-056 ethics-as-coherence). V3 D_V is self-only; V4 social systems extend to D_{V,j} for represented others.

These deferrals are deliberate (see v3_v4_phase_substrate_boundary.md).

Validation experiment (queue after implement-mech269b lock releases)

V3-EXQ-491 (proposed): ARC-054 V3 D_V rollout-horizon validation, 4-arm factorial {use_arc054_dv} x {EXQ-074b condition (z_goal salience contest), EXQ-481 condition (V_s commit-release)}. Acceptance:

  • C1 (EXQ-074b condition, ON arm): z_goal salience competitive with harm salience (goal_action_share / harm_avoid_action_share >= 0.30 mean across 3 seeds; current baseline ~0.0).
  • C2 (EXQ-481 condition, ON arm): freeze-recommit cycle count drops by >=50% vs OFF baseline (current ON_OFF mean = 5.3 releases / 12 recommits per release).
  • C3 (bit-identical OFF): with master flag OFF, all metrics within 1e-9 of pre-flag baseline – enforced by contracts test as standard for V3 substrate landings.
  • C4 (instrumentation sanity): arc054_dv_score non-zero in all 3 seeds when ON; mean(D_V_pi) over evaluated rollouts in the expected range [0.0, 1.0].

Failure on C1 OR C2 with C3 + C4 PASS = V3 D_V form is a true-negative; consider whether the V4 phase form is required (see Section “What V3 does NOT include”). Failure on C3 or C4 = implementation bug, fix and retry.

Lit-pull (deferred until promotion is committed)

Tagged targeted_review_arc054_dv_rollout_horizon/. Targets: rollout-horizon evaluation in hippocampal replay (Pfeiffer & Foster 2013, Mattar & Daw 2018), persistence-of-coupling signals in long-horizon planning (Schuck & Niv 2019), distinction between instantaneous and integrated belief signals in non-stationary environments (Yu & Dayan 2005). Deferred until claim status flip lands cleanly (avoid lit-pull-then-revert churn).

Cross-references

  • V_s invalidation runtime (v_s_invalidation_runtime.md): MECH-269 / MECH-284 / MECH-287 substrate that ARC-054 V3 reads from.
  • V3/V4 phase substrate boundary (v3_v4_phase_substrate_boundary.md): the explicit architectural commitment that V3 working-model uses synaptic V_s + EMA D_V, and that ARC-053 / MECH-225 / MECH-226 / MECH-228 stay v4 as refinements.
  • Verisimilitude intake (evidence/planning/thought_intake_2026-04-09_verisimilitude_ethics.md) Thread A: original formal definition of V(t), D_V, and the TCL.
  • MECH-269b (mech_269b_vs_rollout_gating.md): the parallel session’s symmetric V_s gating on E1/E2 forward predictions. ARC-054 V3 reads downstream of MECH-269b’s gated predictions; the two land as compatible substrate refinements with the same OFF-default contract.

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