MECH-313: stochastic_noise_floor (LC-NE tonic / SAC max-entropy analog)

Claim ID: MECH-313 Subject: policy.stochastic_noise_floor_lc_ne_tonic_analog Status: IMPLEMENTED 2026-05-10 Registered: 2026-05-10 Parent cluster: ARC-065 behavioral_diversity_generation_pathway Sibling claim: MECH-314 structured_curiosity_bonus (separate substrate; not implemented in the same session) Related: MECH-260 (state-dependent dACC anti-recency; Q-045 falsifies whether they collapse) Depends on: ARC-065 (parent), MECH-104 (LC-NE phasic spike, already implemented; MECH-313 is the tonic complement) Blocks: Q-043 weight calibration sweep, Q-045 4-arm collapse ablation (MECH-313 vs MECH-260)

Problem

EXQ-433/433a/433b/470/523/523a/523b were reclassified non_contributory because monomodal policy could not generate balanced agent-vs-env event distributions for the SD-029 C2/C3 measurement gate. EXQ-543 (ARC-062 Phase 2a falsifier) was reclassified non_contributory 2026-05-10 for the same family of reasons. The proposed ARC-064 cluster (bottom-up rule discovery) presupposes behavioural diversity exists in the trajectory record; ARC-062 (top-down rule application) presupposes context to detect. Without an upstream behavioural diversity generator, neither pathway has anything to weight or cluster.

The Pull 1 lit-pull (evidence/literature/targeted_review_arc_065_behavioral_diversity_generation/SYNTHESIS.md, 9 entries, lit_conf 0.78-0.82) settled four falsifiable questions and identified TWO load-bearing missing substrates:

  • MECH-313 stochastic_noise_floor (this claim) – LC-NE tonic analog. State-independent softmax-temperature lift. Wilson 2014 / Faisal 2008 / Aston-Jones & Cohen 2005 anchors.
  • MECH-314 structured_curiosity_bonus (sibling claim, separate substrate, not in this session) – frontopolar exploration analog. Daw 2006 / Friston 2015 anchors.

Pull 1 R1 verdict (BOTH-CHANNELS-NEEDED, conf 0.85): single-channel readings fail empirically (noise-only) or biologically (curiosity-only). Pull 1 R2 verdict assigned LC-NE tonic to LOAD-BEARING. Pull 1 R4 verdict assigned MECH-313 to continuous computation (every tick, regardless of context).

Solution

Module

ree-v3/ree_core/policy/noise_floor.py – NoiseFloor class + NoiseFloorConfig dataclass. Pure-arithmetic, no learned parameters, no nn.Module inheritance. Lives in ree_core.policy.__init__ alongside GatedPolicy (the ARC-062 weak-reading rule-apprehension sibling, also a policy-layer module).

Algorithm

Per waking tick:

effective_T = max(baseline_T + noise_floor_alpha, noise_floor_min_temperature)
  • noise_floor_alpha: SAC-entropy-bonus analog (Haarnoja et al. 2018 soft actor-critic temperature). Constant additive lift on the softmax temperature.
  • noise_floor_min_temperature: hard floor on the effective temperature so the policy never collapses to argmax even under annealing schedules that drive the baseline below 1.0.

Both lift the softmax temperature only. There is no candidate-level bias and no perturbation of scores. The action-selection mechanism is unchanged; the regulator is a pure scalar transform.

Config

REEConfig (ree-v3/ree_core/utils/config.py):

Param Type Default Purpose
use_noise_floor bool False master switch (no-op default)
noise_floor_alpha float 0.1 SAC-entropy-bonus analog; additive temperature lift
noise_floor_min_temperature float 1.0 hard floor; matches existing E3 baseline

All wired through REEConfig.from_dims(). Defaults are conservative starting points; Q-043 calibrates magnitudes via parametric sweep on V3-EXQ-543b/c.

Data flow 

REEAgent.select_action()
  -> noise_floor.compute_effective_temperature(temperature, simulation_mode=False)
  -> e3.select(candidates, effective_temperature, ...)

The integration point is the e3.select(...) call site in select_action(). Phase 1 wires the regulator at the call site only; no candidate-level changes, no E3 internal modification.

Backward compat

  • Master flag default OFF -> agent.noise_floor is None -> select_action passes temperature unchanged to e3.select -> bit-identical to baseline.
  • All existing experiments unaffected.

Phased training

Not applicable. Pure scalar regulator. No learned parameters. No gradient flow.

MECH-094

compute_effective_temperature(simulation_mode=True) returns the baseline temperature unchanged and increments only the simulation-skip counter. Match the SD-035 amygdala / MECH-279 PAG / gated_policy simulation_mode pattern. The select_action call site passes simulation_mode=False (waking action selection); replay / DMN consumers (none today; reserved for forward-compat) get the unmodified baseline so simulation paths cannot inherit the waking-tonic noise floor that biologically belongs only to active behaviour.

Distinction from MECH-260

Aspect MECH-260 (dACC anti-recency) MECH-313 (stochastic noise floor)
Site Per-candidate score_bias Softmax temperature
State State-dependent (recent action history FIFO) State-independent (no history read)
Mechanism Penalises recently-selected action class Lifts entropy uniformly
Biology dACC + lateral aPFC anti-recency suppression (Scholl & Klein-Flugge 2018) LC-NE tonic firing (Aston-Jones & Cohen 2005)
After argmax-on-A behaviour Penalises class A specifically Raises probability mass on every NON-A class equally
ML analog Anti-recency / novelty bonus SAC max-entropy regularisation (Haarnoja 2018)

Q-045 falsifies whether they collapse into a single substrate via a 4-arm ablation (both-OFF / 313-only / 260-only / both-ON). The substrate guarantees they can coexist as independent flags.

Lit-pull verdicts

Pull 1 SYNTHESIS (evidence/literature/targeted_review_arc_065_behavioral_diversity_generation/SYNTHESIS.md, 9 entries, lit_conf 0.78-0.82, supports-direction):

  • R1 BOTH-CHANNELS-NEEDED (conf 0.85): noise-floor (this claim) AND structured-curiosity (MECH-314) both required. Wilson 2014 Horizon task + Faisal/Selen/Wolpert 2008 noise-substrate irreducibility + Friston 2015 active-inference complementary terms. Single-channel readings fail.
  • R2 LC-NE tonic LOAD-BEARING (conf 0.84): Aston-Jones & Cohen 2005 adaptive-gain model. Tonic mode elevates baseline decision noise across all policies. MECH-104 covers the phasic spike (already substrate-landed); MECH-313 is the tonic complement.
  • R3 PROMOTE-TO-CLUSTER (conf 0.82): MECH-313 cannot fold into ARC-051 wanting hierarchy (anatomically distinct frontopolar+IPS vs striatum+vmPFC pathways per Daw 2006) or MECH-260 (which is contributory anti-recency, not active diversity generation).
  • R4 continuous, every tick (conf 0.80): non-zero softmax temperature on E3 every waking tick, regardless of context. Aston-Jones & Cohen 2005 + Friston 2015. NOT triggered.

Magnitudes are NOT settled by the lit-pull. Q-043 captures the open question of relative weight calibration via parametric sweep on V3-EXQ-543b/c.

Validation

Substrate-readiness diagnostic

V3-EXQ-544 (v3_exq_544_mech313_noise_floor_substrate_readiness.py). Five sub-tests UC1-UC5:

  • UC1 forward-pass instantiation
  • UC2 master-OFF backward-compat (agent.noise_floor is None)
  • UC3 effective_temperature lift correctness (alpha + floor arithmetic)
  • UC4 select_action wiring contract (e3 receives the lifted temperature)
  • UC5 MECH-094 simulation gate

experiment_purpose=diagnostic. Substrate readiness only; behavioural validation lands as Q-045 4-arm ablation (MECH-313 OFF / 313 only / 260 only / both ON) on V3-EXQ-543b/c successors AFTER substrate is in.

Behavioural validation (downstream, NOT this session)

  • Q-043 weight calibration sweep on noise_floor_alpha and noise_floor_min_temperature.
  • Q-045 4-arm ablation (MECH-313 vs MECH-260 collapse falsifier).

What this SD enables

  • ARC-065 behavioral_diversity_generation_pathway: completes one of two load-bearing children identified by Pull 1 (the other is MECH-314 structured_curiosity_bonus, a separate substrate task).
  • V3-EXQ-543b/c successors that need behavioural diversity AS INPUT (not as output) – including the rebalanced 4-arm Q-045 ablation experiment.
  • Sleep-substrate refinement experiments (V3-EXQ-418m / 436b / 500b / 503b per the 2026-05-10 reclassification cohort) that need non-degenerate waking diversity for sleep to refine.

Out of scope (deferred)

  • MECH-314 / MECH-314a/b/c structured_curiosity_bonus substrate (separate spawned task; sibling under ARC-065).
  • MECH-318 / MECH-319 substrates (separate spawned tasks).
  • Q-045 4-arm ablation experiment itself (queued AFTER substrate landing).
  • Q-043 parametric sweep (downstream from Q-045 if MECH-313 is shown load-bearing).
  • ARC-065 (parent architectural commitment)
  • MECH-260 (related mechanism; Q-045 falsifies whether they collapse)
  • MECH-104 (LC-NE phasic spike; already substrate-landed; MECH-313 is the tonic complement)
  • MECH-314 (structured_curiosity_bonus sibling under ARC-065; separate substrate)
  • Q-043 (relative weight calibration – parametric sweep)
  • Q-044 (MECH-314a/b/c sub-flavour independence – sibling)
  • Q-045 (MECH-313 vs MECH-260 collapse falsifier – 4-arm ablation)
  • MECH-094 (simulation_mode argument; call-site scoping for waking-only effects)

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