SD-036: GABAergic Cross-Stream Decay Regulator

Claim ID: SD-036 (parent), MECH-279 (PAG freeze-gate child) Subject: gabaergic.cross_stream_decay_regulator Status: candidate, v3_pending Registered: 2026-04-22 Origin exemplar: V3-EXQ-471 seed 0 ep 0 — sustained 200-step avoid-mode lock with z_harm_norm pinned at ~0.7 absent input. See psychiatric_failure_modes.md “Catatonia, Subtype II”. Depends on: SD-010, SD-011, SD-012, MECH-090, MECH-269

Problem

The current V3 substrate has no mechanism by which the absence of input over time returns a latent stream toward baseline. Streams integrate input (e.g. SD-010/SD-011 for harm; MECH-090 for beta) but contain no decay regulator. A single salient event therefore produces a permanently elevated stream value.

V3-EXQ-471 made this concrete: a single hazard contact at t=0 elevated z_harm_norm to ~0.82, which then stayed at ~0.7 for the remaining 199 steps despite zero harm input. Because z_harm was pinned, mode arbitration stayed locked in avoid, the agent froze in place, and energy ran to zero without any homeostatic rescue. The behavioural lock is the surface symptom; the missing decay regulator is the architectural cause.

The architectural commitment SD-036 makes is that decay is not a property of each stream’s update rule (which would distribute the same logic across many modules and produce inconsistent rates) but a property of a regulatory layer that touches multiple streams. The biological analogue is the GABAergic system: a broadly-projecting inhibitory neuromodulatory layer that applies tonic regulation across cortical and subcortical sites simultaneously.

Mechanism

1. Decay as a regulator-layer operation

For each registered stream s ∈ S_decay, the regulator applies in the absence of above-threshold input:

z_s(t+1) = z_s(t) * exp(-tau_s * gaba_tone(t))

where tau_s is a per-stream baseline decay rate and gaba_tone(t) is a global multiplier representing tonic GABAergic level. When stream input exceeds threshold, decay is suspended for that step (the input drives the update). Otherwise, decay proceeds.

The choice of which streams are decay-eligible is an architectural commitment, not a per-stream concern. Initial coverage:

Stream Decay rationale tau (initial)
z_harm_s (sensory harm, SD-010) Single pain event should not drive permanent elevation 0.05 (~20-step half-life)
z_harm_a (affective harm, SD-011) Slower than sensory; emotional residue persists longer 0.02 (~50-step half-life)
z_beta (precision weight, MECH-090) Bistable gate must be able to relax 0.03 (~30-step half-life)
Drive accumulator (SD-012, conditional) Open question — see Open design questions n/a (pending)

2. GABAergic tone as global multiplier

gaba_tone(t) is a slow-varying scalar in [0, 2] representing tonic inhibitory level. At baseline gaba_tone = 1.0 (decay proceeds at the per-stream tau_s). Pharmacological or pathological perturbation maps cleanly:

  • Benzodiazepines / GABA-A agonists: gaba_tone > 1 → faster decay → unlocks frozen streams. This is the architectural prediction matching the clinical observation that benzodiazepines are first-line for catatonia of the SD-036/MECH-279 subtype.
  • GABA-A receptor antibodies / benzodiazepine withdrawal: gaba_tone < 1 → slower decay across all registered streams simultaneously. Predicts the clinical cluster (catatonia + autonomic dysregulation + perceptual disturbance + sleep disruption) as a coherent architectural failure rather than four separate disorders.
  • Chronic stress / glucocorticoid-mediated GABA suppression: sustained gaba_tone < 1 produces the substrate-level vulnerability for harm-stream lock-in. Matches the clinical observation that stress-vulnerable individuals show prolonged threat responses.

3. PAG freeze-gating (MECH-279)

The freeze response is a committed behavioural state — sustained motor immobility plus elevated autonomic arousal. It is not just “no movement”; it is an active commitment to not-move that itself has a duration and an exit criterion. Biologically this is gated by the periaqueductal gray (PAG), where descending inputs from amygdala, hypothalamus, and medial PFC converge on a population whose activity determines freeze vs flight vs fight selection. PAG freeze-promoting cells are themselves GABAergic, and freeze termination requires GABAergic inhibitory control to wane.

MECH-279 commits to this architecture:

freeze_commit(t) = (z_harm_a(t) * duration_above_threshold(t)) > theta_freeze
freeze_active(t) = freeze_commit OR (freeze_active(t-1) AND z_harm_a(t) > exit_threshold)
exit_threshold = theta_freeze * gaba_tone(t)

When freeze is active, the action selector is constrained to no-op / minimal-movement actions. Exit requires z_harm_a to fall below exit_threshold, which depends on both SD-036 decay (z_harm_a returns toward baseline) and gaba_tone (GABA agonists raise the exit_threshold, making exit easier). This means the same neurotransmitter system gates both the entry and the exit — entry via PAG freeze-cell commitment, exit via SD-036 decay returning z_harm_a below the exit threshold. The clinical observation that GABA agonists treat freeze catatonia is an architectural prediction, not an empirical add-on.

Why a regulator layer rather than per-stream decay

Two reasons:

  1. Coordination across streams. Pathological GABAergic states (withdrawal, antibody syndromes) produce clusters of failures, not single-stream failures. Putting decay on a regulator layer means a single perturbation (gaba_tone < 1) explains the cluster. Per-stream decay would require coordinated multi-stream perturbation to produce the same clinical picture, which is architecturally less parsimonious and clinically harder to motivate.

  2. Pharmacological tractability. Benzodiazepines, GABA-A antagonists, alcohol, and barbiturates all act at the GABA-A receptor. A regulator-layer model means these compounds have a single point of action with broad downstream effects, matching their pharmacology. A per-stream decay model would predict that benzodiazepines should have selective effects on specific streams — which is not what is observed.

The same logic applied to give the cingulate substrate (SD-032) a network-level coordinator (SD-032a) rather than per-subdivision coordination. The principle: multi-target regulation lives at the regulator, not in each target.

Relationship to existing claims

Existing claim Relationship
SD-010 (z_harm_s sensory harm stream) Provides the substrate stream; SD-036 provides decay back to baseline. SD-010 was registered without decay; this is a substrate completion, not a contradiction.
SD-011 (z_harm_a affective harm stream) Same as SD-010.
SD-012 (drive-modulated goal seeding) Drive accumulator is candidate for SD-036 coverage. Open question whether drive should decay (vs. the homeostatic-override mechanism providing its own dynamics — see homeostatic_override_litpull).
MECH-090 (bistable beta gate) Bistability is preserved; SD-036 provides the decay channel that allows the gate to relax out of a stuck high-beta state. EXQ-471 showed z_beta_val flat at 0.0019 across 200 steps — consistent with no decay channel.
MECH-269 (hippocampal anchor selection / regional verisimilitude) Anchor reset in MECH-269 should be informed by SD-036 decay state — when the harm stream has decayed, the anchor based on the original harm event becomes invalid. See hippocampal_anchor_selection.md post-EXQ-471 update.
MECH-186/187/188 (serotonergic V4 mechanisms) Different neurotransmitter system, different timescale. 5-HT modulates terrain/transduction/maintenance for the goal pipeline; GABA modulates decay across streams. They are co-operating regulatory layers, not alternatives.

Predicted observables (V3 scope)

A V3 experiment validating SD-036 + MECH-279 would measure:

  1. EXQ-471 lock pattern resolution. Re-running V3-EXQ-471 with SD-036 enabled (default tau_harm_s = 0.05, gaba_tone = 1.0) should produce mode flip from avoid back to a goal-seeking mode by ~t=50 (when z_harm_norm decays below the mode-arbitration threshold). The seed-0 ep-0 trace would shift from “200 steps stuck” to “~50 steps avoid, then resumed exploration.”

  2. GABA tone manipulation. Setting gaba_tone = 1.5 (benzo-analog) should accelerate recovery to ~30 steps; gaba_tone = 0.5 (withdrawal-analog) should extend lock to ~100 steps or beyond. The dose-response curve is the falsifiable prediction.

  3. Multi-stream cluster. Setting gaba_tone = 0.3 should produce simultaneous failures across z_harm_s, z_harm_a, and z_beta — concretely, the agent should show prolonged harm responses, persistent over-precision, and freeze-prone behaviour together, not in isolation. This is the architectural signature distinguishing SD-036 from per-stream decay alternatives.

  4. MECH-279 freeze commit/exit. Inducing sustained z_harm_a above theta_freeze for >5 steps should produce committed freeze (action repertoire constrained); exit should require z_harm_a < exit_threshold AND should be accelerated by gaba_tone > 1. Catatonia-equivalent (gaba_tone = 0.7, sustained harm) should produce sustained freeze with elevated z_harm_a, matching the EXQ-471 phenotype.

Candidate experiment names (not yet queued):

  • v3_exq_NNN_sd036_decay_unlocks_exq471.py (single-condition: SD-036 on/off, replicate EXQ-471 conditions)
  • v3_exq_NNN_sd036_gaba_tone_dose_response.py (multi-condition: gaba_tone ∈ {0.3, 0.5, 1.0, 1.5, 2.0})
  • v3_exq_NNN_mech279_pag_freeze_gate.py (sustained-threat protocol; measure freeze commit, exit, and gaba_tone modulation)

Substrate hooks required:

  • ree_core/regulators/gabaergic_decay.py — new module with GABAergicDecayRegulator registering streams and applying per-step decay.
  • ree_core/pag/freeze_gate.py — new module hosting MECH-279 freeze commit/exit logic; consumes z_harm_a and gaba_tone; emits freeze_active boolean to action selector.
  • Wiring in ree_core/agent.py to tick the regulator after stream updates and before mode arbitration.

Open design questions

  1. Drive accumulator decay. SD-012’s drive accumulator (energy / hunger) is candidate for SD-036 coverage but ambiguously so: drive should keep accumulating under sustained deprivation (that is its point), so a simple decay would defeat its function. The likely resolution is that drive is not covered by SD-036 — instead, the homeostatic-override mechanism (separate, V4-or-late-V3) provides the dynamics that allow drive to outweigh threat under survival demand. SD-036 then covers regulatory streams (harm, beta, possibly precision) but not homeostatic accumulators. Defer to homeostatic_override_litpull synthesis.

  2. gaba_tone source. In V3 baseline, gaba_tone = 1.0 constant is the simplest first pass. A more complete model would have gaba_tone driven by sleep state (elevated during NREM, reduced during REM), by stress accumulation (chronic stress suppresses tone), or by external pharmacological state (for medication modeling). Defer specification to a V4-style 5-HT/GABA cross-state architecture (parallel to MECH-203/204 for serotonin).

  3. Per-stream tau learning. First pass uses fixed per-stream tau_s. Long-run, the decay rate could itself be plastic — e.g. an agent with chronic harm exposure might slow tau_harm_s (architectural model of stress-induced sensitization). Open question whether this belongs in SD-036 or in a downstream plasticity claim.

  4. Interaction with the bistable beta gate (MECH-090). MECH-090 was designed as bistable but EXQ-471 showed it flat — either the bistability never engaged or the stuck-state has no exit. Adding SD-036 decay to z_beta may resolve this by providing the exit channel. Worth confirming experimentally that SD-036 + MECH-090 together produce bistable-with-relaxation behaviour, not just slow-monotonic behaviour.

  5. Failure mode to surface as diagnostic. If gaba_tone is set too high, harm response is suppressed too rapidly and the agent fails to learn from harm events (no time for residue field update). Architecturally this matches the clinical observation that benzodiazepine-treated patients show impaired fear-learning. Worth surfacing as a diagnostic counter rather than relying on the dose-response sweep to catch it.

Status log

  • 2026-04-22 — Design doc written. SD-036 and MECH-279 reserved. Discussion origin: V3-EXQ-471 fishtank visualization (user observation 2026-04-22) showed 200-step catatonic lock with pinned z_harm_norm despite zero harm input; user proposed GABAergic decay as the missing regulator; agreement that catatonia subtype II is architecturally distinct from MECH-202B commit-gate paralysis.
  • Registration in claims.yaml follows in same session.

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