Threshold Supervisor: Scattered Mechanisms Survey

Registered: 2026-05-07 Status: research-anchor survey supporting Q-041

Purpose

REE has accumulated several adaptive-threshold and adaptive-baseline mechanisms, each scoped to one substrate. The brain dynamically manages thresholds for commitment, release, surprise, urgency, and pass/fail across multiple timescales – and this is a stability lever, not just a tuning knob. This document inventories what is already adaptive in REE, the timescale each operates on, and what a unified meta-level “threshold supervisor” would have to add that is not already covered.

This is a registration anchor for Q-041. It is not a design doc. The substrate question is what the experiments referenced from Q-041 are meant to resolve.

Existing scattered mechanisms

Locus Adaptive quantity Update timescale Read by Notes
ARC-016 (E3 dynamic precision) precision = f(running variance of E3 prediction error) per-step EMA, ~10-100 steps E3 score weighting; commit threshold via relative ratio Internal to E3. Does not adapt cross-substrate thresholds. EXQ-018b PASS validated the variance-driven relative-threshold path; EXQ-396a/b and EXQ-454 reclassified non_contributory under the V_s-monostrategy substrate gap.
MECH-258 / SD-032b (dACC pe_cap, FIFO action history) precision-weighted PE normaliser, recency action-history window window = 8 actions (dacc_suppression_memory) dACC bundle -> E3 score_bias Partial volatility tracking. NOT a Behrens-2007-style explicit volatility estimator; the pe_cap is a fixed scale parameter, not learned.
SD-032c (AIC-analog interoceptive baseline EMA) drive/fatigue baseline against which z_harm_a salience is computed EMA, alpha set per AICConfig aic_salience, harm_s_gain Floats with z_harm_a – urgency is computed against a moving baseline rather than a fixed threshold. Drive-dependent by construction.
SD-032d (PCC-analog stability scalar) scalar in [0,1] from success EMA + drive_level + steps-since-last-offline success_alpha=0.02 (~50-step half-life); offline_recency_window=500 SalienceCoordinator effective_threshold Modulates the SD-032a switch threshold. Ties offline cadence to threshold sensitivity.
SD-032e (pACC drive bias EMA) bounded drive_bias added to SD-012 drive_level alpha=0.002 (~347-step half-life) GoalState, SalienceCoordinator, AIC, PCC, dACC bundle Slowest of the cingulate cluster. The meaning of “drive=0.7” drifts under chronic load. Closest existing analogue to neuromodulator-setpoint chronic shift.
MECH-040 (safety baseline / volatility) dual control channels (provisional, no genuine evidence) unspecified control_plane Registered but never validated on real substrate.
MECH-204 (REM zero-point capture) precision_at_rem_entry snapshot for sleep recalibration per-sleep-bout (capture-only; no writeback) Captures the recalibration target but does not act on it. The sleep-side writeback is unimplemented – this is the most obvious gap.

Timescale spread is wide – ~10 steps (ARC-016 EMA) up to ~350 steps (SD-032e pACC drive bias) – but no mechanism operates at the cross-substrate / sleep-bout / multi-day timescale at which neuromodulator setpoints recalibrate biologically.

What is missing

A meta-level threshold supervisor would adapt pass/fail or commit/release thresholds across substrates based on system-wide instability metrics. Specifically:

  1. Cross-substrate volatility tracking. Each adaptive locus above tracks variance internal to its own signal. Nothing aggregates across loci. A genuine volatility estimator (Behrens et al 2007, Nat Neurosci) would read multiple PE streams jointly and emit a system-level learning-rate signal that downstream thresholds consume.

  2. Sleep-mediated writeback. MECH-204 captures a precision zero-point at REM entry but the writeback path that uses it to recalibrate waking thresholds is absent. The brain treats sleep as global recalibration (Tononi & Cirelli 2014, Neuron, SHY); REE captures the snapshot and discards it.

  3. Setpoint drift under chronic load. SD-032e’s pACC drive bias is the closest existing analogue but is scoped to a single substrate (drive_level write-back from z_harm_a). Chronic-stress 5-HT/DA setpoint shifts in biology recalibrate many downstream thresholds simultaneously (commit, release, surprise) – REE has no joint mechanism.

  4. Coherence guarantee. Independent EMAs at different timescales can drift into incoherent regimes (e.g., high pACC drive_bias with low PCC stability scalar). A supervisor would either enforce a coherence constraint or expose the divergence as an instability signal in its own right.

Why this is a question, not a substrate

The natural temptation is to write this as a SD (“SD-NNN: threshold supervisor module that reads all adaptive loci and emits joint recalibration signals”). I have resisted that for two reasons:

  • Premature commitment. Whether the supervisor is needed at all is an open empirical question. The scattered mechanisms may already produce coherent meta-stability under realistic load – the V_s-monostrategy substrate gap currently masks the dependent behaviour, so we do not yet know.
  • V3-vs-V4 placement is genuinely uncertain. The simplest threshold supervisor is a slow EMA of EMAs – could land in V3. The full sleep-mediated writeback supervisor requires MECH-204 to act on its snapshot, which depends on the sleep substrate enrichment (MECH-285, MECH-286, INV-049 implementation). That is a V4 commitment.

Q-041 registers the question. If experimental evidence licenses a substrate-level commitment, the cluster (SD + supporting MECH) follows.

Anchor literature (registration-time, pre-lit-pull)

  • Behrens et al. 2007. Learning the value of information in an uncertain world. Nat Neurosci 10:1214-1221. – dACC tracks volatility and adjusts learning rate accordingly.
  • Friston & Adams 2013 / Adams, Shipp & Friston 2013. Predictions not commands: active inference in the motor system. – precision-weighted PE as the canonical adaptive-control variable.
  • Tononi & Cirelli 2014. Sleep and the price of plasticity. Neuron 81:12-34. – SHY: sleep as global synaptic recalibration. Bears directly on the MECH-204 writeback gap.

A targeted lit-pull on “adaptive learning rate dACC volatility” + “neuromodulator setpoint chronic stress recalibration” + “synaptic homeostasis SHY recalibration” is the natural Q-041 successor; staged here as a follow-on rather than gating the registration.

Observable signature for a working supervisor

Under sustained drive_level=0.9 for ~1000 steps, an agent equipped with a meta-level threshold supervisor should show coherent shift across substrates – effective commit threshold, effective beta-gate release threshold, AIC switch threshold, dACC pe_cap normalisation should move together along a single low-dimensional trajectory consistent with a shared volatility/setpoint signal. Without a supervisor, the adaptive loci move independently – the trajectory is high-dimensional and substrate- specific.

This is the signature the Q-041 diagnostic experiment proposal targets.

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