MECH-319: Simulation-Mode Rule-Write Gate (Categorical Replay Tag)

Claim ID: MECH-319 Subject: policy.arbitration.simulation_mode_write_gating_substrate_ree_novel_function Status: IMPLEMENTED 2026-05-10 (substrate landed; behavioural validation deferred to V3-EXQ-543c-successor) Registered: 2026-05-10 (cluster registration session) Depends on: MECH-094 (architectural principle), MECH-312 (arbitration layer) Blocks: V3-EXQ-543c-successor (artificial-write-channel-routing falsifier)

Problem

MECH-094 names the architectural principle that simulation / replay / DMN-generated content must NOT update the agent’s persistent rule state, by being routed through a categorical write gate keyed to a hypothesis tag. The principle is silent on which substrate carries the categorical signal and at which layer the gate operates.

In the V3 substrate prior to this commit, three arbitration-layer write sites each handled simulation gating in isolation:

  • GatedPolicy.forward(simulation_mode=True) (ARC-062 Phase 1) returns zeros and increments a skip counter.
  • LateralPFCAnalog.update(...) (SD-033a) is mode-conditioned via MECH-261 write_gate("sd_033a") (continuous mode gate, not a categorical simulation-mode tag).
  • HippocampalModule ghost-goal probes (MECH-293) carry hypothesis_tag=True per record_committed_trajectory.

Three problems:

  1. No unified seam. Each call site has its own simulation-mode handling, with no central authority that translates a caller- supplied tag and a debug / falsifier flag into the final admit/block decision.
  2. No diagnostic surface. Experiments cannot observe how often the gate fires, at which sites, and under which mode (waking vs replay).
  3. No falsifier pathway. Phase 2 of arc_062 requires a config flag that artificially routes simulation-mode passes through the write channel anyway, predicting monomodal-collapse re-emergence via the MECH-094 / MECH-319 generalisation. Without a single coordinator, the V3-EXQ-543c falsifier would have to flip multiple per-module flags inconsistently.

MECH-319 is the substrate-level instantiation of MECH-094 at the rule- arbitration layer that solves all three.

Substrate-availability premise

Well-anchored biologically (Pull 3 SYNTHESIS R1 GENUINE-NOVELTY-CONFIRMED conf 0.72; literature in evidence/literature/targeted_review_mech_312_arbitration_divergences/):

  • Joo & Frank 2018 (Nat Rev Neurosci): SWR review. The discriminable sharp-wave-ripple machinery is the biological substrate that carries the categorical “this is replay, not waking” signal. Treats downstream effects as continuous strength modulation; the write-gate function is REE-specific.
  • Foster & Wilson 2006 (Nature): reverse replay. Concrete structural marker (temporal-reverse compressed sequences during awake quiescence) demonstrating the substrate-level identifiability that MECH-319’s categorical tag relies on. Documents the discriminable signal but does not show downstream regions exploiting it as a write-gate.

REE-novel function

The architectural commitment that downstream rule-write machinery uses the discriminable replay signature to suppress arbitration-weight updates that would otherwise fire if the same rule were applied during waking action. Plausible at the cellular / synaptic level but biologically unverified (Pull 3 R1 verdict). The V3 falsification path is the artificial-write-channel-routing config flag (admit_writes=True) threaded through this module’s effective_simulation_mode primitive.

Solution

Module

ree-v3/ree_core/regulators/simulation_mode_rule_gate.py – pure- arithmetic regulator (no nn.Module inheritance, no learned parameters) sibling to GABAergicDecayRegulator (SD-036) and BroadcastOverrideRegulator (SD-037).

Single primitive 

gate.effective_simulation_mode(simulation_mode: bool, site: str) -> bool

Translates the caller-supplied simulation tag and the configured falsifier flag into the final admit/block decision the downstream arbitration-write site should observe.

Truth table (master_on, admit_writes, caller_sim) -> output:

master admit_writes caller_sim output semantic
OFF (any) (any) identity (caller_sim) bit-identical pre-MECH-319
ON False False (waking) False admit waking write
ON False True (replay) True block simulation write – MECH-319 normal
ON True False (waking) False admit waking; flag has no effect
ON True True (replay) False admit simulation write – V3-EXQ-543c falsifier

The gate is idempotent for waking calls (always returns False), so wiring it into existing waking call sites is bit-identical regardless of admit_writes. The falsifier-control asymmetry surfaces only when caller_sim=True (replay paths, ghost-goal probes, DMN passes).

Diagnostic counters

Per-call + per-site:

  • n_calls_total, n_waking_admitted, n_simulation_blocked, n_simulation_admitted (the falsifier path).
  • per_site_calls, per_site_waking_admitted, per_site_simulation_blocked, per_site_simulation_admitted – keyed on canonical site labels SITE_GATED_POLICY, SITE_LATERAL_PFC, SITE_DEFAULT. New consumer call sites can pass arbitrary strings.

Config

  • REEConfig.use_simulation_mode_rule_gate: bool = False – master, bit-identical OFF.
  • REEConfig.simulation_mode_rule_gate_admit_writes: bool = False – V3-EXQ-543c falsifier flag.

Construction raises ValueError when admit_writes=True without the master flag also on (loud-not-silent guard against mis-configuration – admit_writes is meaningless without the substrate to gate).

Agent wiring

Two existing arbitration-write call sites in REEAgent.select_action consult the gate when it is instantiated:

  1. GatedPolicy block (line ~2752): the literal simulation_mode=False passed to gated_policy.forward(...) is replaced by gate.effective_simulation_mode(False, site=SITE_GATED_POLICY). Bit-identical for the waking path; the seam is exposed for V3-EXQ-543c replay-driven invocations.

  2. LateralPFCAnalog block (line ~2640): consult the gate before lateral_pfc.update(...).

    eff_sim = gate.effective_simulation_mode(False, site=SITE_LATERAL_PFC)
if eff_sim:
    skip lateral_pfc.update(...)
else:
    proceed with the existing MECH-261 mode-conditioned EMA

    The compute_bias call still runs (arbitration RECEIVES the bias even during simulation – it just does not write back into rule_state on simulation ticks). Bit-identical for waking; falsifier-routed simulation writes to lateral_pfc would be skipped under default MECH-319 ON, admitted under admit_writes=True.

REEAgent.reset() calls gate.reset() per episode to clear diagnostic counters. The gate has no persistent state across ticks.

MECH-094 invariance

This substrate does NOT modify MECH-094, GatedPolicy.forward’s simulation_mode argument semantics, or LateralPFCAnalog.update. The gate is a pre-call coordinator that wraps the simulation_mode argument that callers ALREADY pass. With MECH-319 disabled, every arbitration-write call site behaves bit-identically to its pre-MECH-319 form.

Backward compatibility

  • use_simulation_mode_rule_gate=False (default) -> agent.simulation_mode_rule_gate is None; both call sites take the legacy literal path.
  • 288/288 contract + preflight tests PASS with master OFF (regression-clean; the suite was 273 pre-MECH-319, plus 15 new MECH-319 contracts).
  • With master ON and waking-only callers (the only regime select_action currently exercises), the gate returns False at every site. Bit-identical to baseline.

Architecture Context

Relation to MECH-094

KEEP-AS-IS verdict (Pull 3 R1 + Pull 4 R3): MECH-094 is not edited.

MECH-094 names the architectural principle: a categorical phi(z) write gate keyed to a hypothesis tag, with the simulation-mode- suppressed write channel. MECH-319 names the specific substrate instantiation: the unified categorical write gate that operates at the rule-arbitration layer (MECH-312 sub-mechanisms), with SWR machinery as the substrate-level analog of the simulation tag.

The two coexist:

  • MECH-094 is the principle that downstream architectural work may instantiate at multiple layers (rule arbitration, rule-state abstraction, future arbitrators).
  • MECH-319 is the first such substrate-level instantiation, scoped to the rule-arbitration layer.

Future architectural work may name additional MECH-319-siblings for other function-sites where MECH-094’s write-gate principle should apply (e.g. MECH-094 at the rule-state abstraction layer of MECH-318). These are deferred until the substrate sites are themselves wired.

Relation to MECH-261

MECH-261 (mode-conditioned write-gate registry on SD-032a SalienceCoordinator) is a complementary, continuous gate. MECH-261 returns a per-mode weight in [0, 1] that scales the magnitude of the EMA update on LateralPFCAnalog.rule_state. MECH-319 is a categorical (binary admit/block) pre-gate keyed to the simulation tag of the caller, not the operating mode of the agent.

The two gates compose:

  • Caller is waking (caller_sim=False) -> MECH-319 admits -> MECH-261 modulates the EMA strength based on operating mode.
  • Caller is simulation (caller_sim=True), MECH-319 normal -> MECH-319 blocks the entire update() call, MECH-261 is never consulted.
  • Caller is simulation, MECH-319 falsifier (admit_writes=True) -> MECH-319 admits -> MECH-261 modulates as if the caller were waking.

Phase 1 vs Phase 2

This is a Phase 1 substrate landing. The behavioural test that flips admit_writes=True and routes a replay-driven invocation through the rule-arbitration layer is V3-EXQ-543c-successor, deferred until the MECH-313 / MECH-314 / MECH-318 sibling substrates have landed (see arc_062_rule_apprehension_plan.md GAP-K and the 2026-05-10 cluster-registration decision-log entry).

What This SD Enables

  • V3-EXQ-543c-successor falsifier: a paired-arm experiment comparing admit_writes=False (MECH-319 normal) vs admit_writes=True (artificial-write-channel-routing) under a replay-driven invocation path. Predicted: monomodal-collapse re-emergence under admit_writes=True. This experiment validates BOTH MECH-094 and MECH-319 simultaneously.
  • Future MECH-319-siblings at other arbitration sites can plug into the same gate by passing a new site label.
  • Diagnostic observability: experiment manifests can include gate.get_state() to report how often the gate fired and at which sites.

Validation experiment

V3-EXQ-546 substrate-readiness diagnostic queued. Six sub-tests (UC1-UC5 + UC3b precondition):

  • UC1 forward-pass instantiation + canonical diagnostic keys.
  • UC2 master-OFF backward-compat (agent.simulation_mode_rule_gate is None).
  • UC3 truth-table coverage across the 6 valid (master, admit_writes, caller_sim) combinations.
  • UC3b precondition raises ValueError on admit_writes=True without master ON (loud-not-silent).
  • UC4 select_action wiring contract: gate sees waking calls from both gated_policy and lateral_pfc sites after one act_with_split_obs tick; n_simulation_* counters remain zero.
  • UC5 MECH-094 invariance: master-OFF and master-ON-with-waking- caller produce bit-identical wiring outputs; asymmetry surfaces only at caller_sim=True.

Smoke 6/6 PASS 2026-05-10. 15 contract tests in tests/contracts/test_mech_319_simulation_mode_rule_gate.py PASS. 288/288 full preflight + contracts PASS.

Behavioural validation – the V3-EXQ-543c-successor falsifier with the admit_writes=True arm and a replay-driven invocation path – is queued separately AFTER MECH-313 / MECH-314 / MECH-318 sibling substrates have landed.

  • MECH-094 (architectural principle this substrate instantiates; KEEP-AS-IS per Pull 3 R1 + Pull 4 R3 verdicts).
  • MECH-312 (the arbitration layer whose write-gate this implements; parent of the MECH-312a/b/c/d sub-mechanisms that the gate suppresses during simulation passes).
  • MECH-261 (mode-conditioned continuous write gate; complementary to MECH-319’s categorical pre-gate).
  • ARC-062 (rule apprehension; arc_062 GAP-K hosted this substrate landing).
  • MECH-313 / MECH-314 / MECH-318 (sibling ARC-065 / ARC-064 substrates landed in separate spawned tasks the same day).
  • MECH-293 (ghost-goal probes; future call site that will pass simulation_mode=True through the gate).
  • SD-033a (LateralPFCAnalog; one of the two wired arbitration-write call sites).
  • MECH-309 (logical-necessity claim that motivates the broader ARC-062 cluster).

Lit-pull anchors

  • evidence/literature/targeted_review_mech_312_arbitration_divergences/synthesis.md R1 GENUINE-NOVELTY-CONFIRMED verdict (conf 0.72) + R5 KEEP-AS-IS recommendation for MECH-094 paired with MECH-319 registration as REE-novel candidate substrate claim.
  • Pull 4 R3 KEEP-AS-IS reaffirmation.

Files touched in landing pass (2026-05-10)

  • ree-v3/ree_core/regulators/simulation_mode_rule_gate.py (NEW)
  • ree-v3/ree_core/regulators/__init__.py (export)
  • ree-v3/ree_core/utils/config.py (REEConfig fields + from_dims kwargs)
  • ree-v3/ree_core/agent.py (import + __init__ instantiation + select_action GatedPolicy + LateralPFC site wiring + reset hook)
  • ree-v3/tests/contracts/test_mech_319_simulation_mode_rule_gate.py (NEW, 15 tests)
  • ree-v3/experiments/v3_exq_546_mech319_simulation_mode_rule_gate_substrate_readiness.py (NEW)
  • ree-v3/experiment_queue.json (V3-EXQ-546 appended)
  • ree-v3/CLAUDE.md (MECH-319 SD entry appended)
  • REE_assembly/docs/architecture/mech_319_simulation_mode_rule_gate.md (THIS FILE, NEW)
  • REE_assembly/docs/claims/claims.yaml (MECH-319 status + evidence_quality_note update)
  • REE_assembly/evidence/planning/arc_062_rule_apprehension_plan.md (GAP-K row + decision-log entry)

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