MECH-271 Hypothesis-Tag-as-Routing – V3 Substrate Landing Plan

Claim: MECH-271 (hypothesis tag realised as differential downstream routing of hippocampal replay, not source-side boolean). Status: candidate, implementation_phase: v3, v3_pending: true. Registered: 2026-04-21. Plan written: 2026-04-26. Depends on: MECH-094 (the tag this claim reframes), MECH-269 (anchor/probe distinction), MECH-269b (parallel-session: symmetric V_s gating on E1/E2 – MECH-271 routing reads from gated streams), ARC-007 (path memory), ARC-018 (viability map), SD-033a (lateral-PFC-analog destination), SD-035 (BLA-analog destination). Architectural commitment (V3): the hypothesis tag is implemented as a routing- destination weighting vector attached to each replay event, not as a boolean write gate. The V3 form uses synaptic destination weights (no ephaptic substrate); V4 will refine the routing signal to a phase-channel coordinate (MECH-228 ephaptic field-coherence).

Why this is “small” (the user’s framing)

Most of the substrate MECH-271 needs is already implemented:

  • MECH-269 anchor/probe distinction – LANDED 2026-04-22 (Phase 2 ii AnchorSet, dual- trace mark_inactive not erase, k=5 hysteresis).
  • MECH-094 hypothesis tag (categorical write gate) – LANDED earlier (REEAgent simulation_mode flag; consumed by MECH-275/273 sleep aggregator paths).
  • SD-033a lateral PFC analog – LANDED 2026-04-20 (V3-EXQ-456 PASS).
  • SD-035 amygdala-analog (BLA + CeA) – LANDED 2026-04-21 (V3-EXQ-473/474 PASS).
  • ARC-018 viability map + ARC-007 path memory – LANDED earlier in V3.

What MECH-271 adds is the routing table + the write-destination dispatcher + the audit hook that flags tag loss. This is roughly 200-300 LOC plus a contracts test suite – consistent with the “small” framing.

V3 substrate (what to land)

1. RoutingTable (new, ree_core/hippocampal/replay_routing.py)

Static destination weights per (replay_kind, destination_name) pair:

destination_name anchored probe source claim
e1_consolidation 1.0 0.0 MECH-271 (subiculum -> entorhinal -> neocortex)
lateral_pfc_viability_map 1.0 0.0 SD-033a (anchored destination)
bla_affective_tagging 0.2 1.0 SD-035 (probe destination)
nac_curiosity_scaffolding 0.0 1.0 NAc novelty fan-out
e3_planning_evaluation 1.0 0.4 both kinds inform planning, anchored dominates

replay_kind is determined at draw time:

  • Anchored = MECH-269 anchor-set draw (active anchor with V_s_anchor >= reset threshold).
  • Probe = MECH-285 sleep-priority draw (high-staleness softmax) OR an explicit is_probe=True waking draw (deferred V4: waking probe replays).

2. ReplayEvent payload extension

Existing replay events already carry simulation_mode (MECH-094 boolean). Add:

  • replay_kind: Literal["anchored", "probe"]
  • routing_weights: dict[str, float] – looked up from RoutingTable at emission time

Both fields are Optional and default to backward-compatible values when MECH-271 is disabled (anchored / {e1_consolidation: 1.0, lateral_pfc_viability_map: 1.0} – the pre-MECH-271 implicit assumption).

3. Destination dispatcher (in REEAgent.consume_replay)

Each downstream consumer (E1 ContextMemory writeback, SD-033a LateralPFCAnalog gate, SD-035 BLA encoding, NAc-curiosity scaffolding, E3Selector) is given a destination_name registration string. The dispatcher passes the replay event to a consumer if routing_weights[destination_name] >= consumer_threshold (default 0.5, configurable per consumer).

This is the only behavioural change. With the master flag OFF, all consumers see all replays (current behaviour); with the master flag ON, anchored consumers see only anchored replays and probe consumers see only probe replays, modulated by the table.

4. Audit hook (MECH-094 tag-loss detector)

A new tag_loss_audit counter increments when a probe replay reaches a destination with anchored_weight > 0.8 AND probe_weight < 0.2. This is the auditable confabulatory- planning signature MECH-271 names. Surface as mech271_tag_loss_count in SleepCycleState.last_metrics + REEAgent.diagnostic_signals.

In a healthy substrate this counter stays at 0; under deliberate ablation (mech271_force_tag_loss_rate=0.X) it accumulates at the expected rate. The audit hook is the substrate-level test that makes confabulation distinguishable from psychosis (the psychosis signature is at MECH-094 source-side; confabulation is the routing-side leak this counter catches).

5. Configuration

Master flag use_mech271_routing (default False, bit-identical-when-OFF preserved). Sub-knobs:

  • mech271_consumer_threshold (default 0.5)
  • mech271_force_tag_loss_rate (default 0.0; ablation knob for the audit-hook test)
  • mech271_routing_table_path (default uses the static table above; path allows experiment-specific overrides without code edit)

Contracts test suite (tests/contracts/test_mech_271_routing.py)

10 tests, parallel to the SD-035 / Sleep-Phase-A pattern:

  • M1 import + module exposure
  • M2 default-OFF backward-compat (ContextMemory, LateralPFCAnalog, BLA all see all replays)
  • M3 OFF-no-instantiation (no RoutingTable created when flag OFF)
  • M4 ON-anchored-replay reaches e1_consolidation + lateral_pfc, NOT bla / nac
  • M5 ON-probe-replay reaches bla + nac, NOT e1_consolidation
  • M6 e3_planning_evaluation receives both kinds (mixed-destination correctness)
  • M7 audit hook fires on forced tag loss (force_tag_loss_rate=1.0 -> count == n_probe)
  • M8 audit hook quiet on healthy substrate (force_tag_loss_rate=0.0 -> count == 0)
  • M9 routing_table_path override loads correctly
  • M10 bit-identical-OFF leaks no mech271_* keys

Validation experiment (queue after implement-mech269b lock releases)

V3-EXQ-492 (proposed): MECH-271 routing-as-tag substrate validation. 4-arm:

  • A0_off: master flag OFF (baseline; reproduces current behaviour)
  • A1_routing_only: master flag ON, force_tag_loss_rate=0.0 (clean substrate)
  • A2_forced_leak: master flag ON, force_tag_loss_rate=0.5 (deliberate confabulation injection)
  • A3_high_leak: master flag ON, force_tag_loss_rate=1.0 (full tag loss)

Acceptance:

  • C1: A1 reduces consolidation writes from probe replays vs A0 (>=70% reduction in ContextMemory write events tagged probe).
  • C2: A2 audit count tracks injection rate within +/-10% (instrumentation correctness).
  • C3: A3 reproduces a confabulation behavioural signature – specifically, an increase in E1-prediction-divergence on held-out probe content (>=2x relative to A1).
  • C4: bit-identical OFF (A0 vs pre-flag baseline within 1e-9 on all run metrics).

The behavioural signature in C3 grounds MECH-271’s clinical claim that probe content reaching consolidation produces an auditable confabulatory-planning failure mode distinct from psychosis (which is a source-side MECH-094 failure – different experiment).

Failure modes / known caveats

Static routing table is a V3 simplification. Biology likely has the routing weights adapt under learning (e.g. cortical schemas that gain reliable predictive value should attract more anchored-replay routing; novel schemas should gain probe-routing capacity). Adaptive routing is V4 work, dependent on MECH-228 ephaptic field-coherence (the biological mechanism that lets routing be dynamically modulated per-event).

Single audit counter is a V3 minimum. A richer audit would track per-destination leak rates and per-source-region leak rates – diagnostic for tracing which schema is driving confabulation. V3 lands the single counter; V4 may add per-region breakdown if clinical use cases (psychiatric simulation experiments) require it.

No phase-channel routing is the V4 deferral. MECH-271’s full claim is that the tag is realised as ephaptically-coordinated routing – the phase channel IS the tag at the biophysical level. V3 approximates this with a discrete routing table; V4 adds the continuous phase-channel coordinate per MECH-228 + the per-destination phase-coupling strength per ARC-053 TCL.

Cross-references

  • MECH-094: source-side tag (already implemented; this claim refines downstream realisation).
  • MECH-269 / MECH-269b (parallel session): anchor/probe distinction at source + symmetric V_s gating on E1/E2.
  • SD-033a, SD-035: anchored / probe destinations whose dispatchers MECH-271 wires.
  • v3_v4_phase_substrate_boundary.md: explicit V4 deferral commitment for the phase-channel realisation.
  • v_s_invalidation_runtime.md: parent V_s document that names MECH-271 as the routing-side companion claim.

Lock coordination

This plan does not edit ree-v3/ree_core/hippocampal/module.py – that file is held by the active implement-mech269b session. MECH-271 substrate code lands in a follow- up /implement-substrate session, scheduled after implement-mech269b releases its lock. That ordering is appropriate substrate-wise too: MECH-271 routing reads from the V_s gating that MECH-269b adds, so MECH-269b lands first.

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