Slow Modulatory State and Compulsive Loops
Status: V4/V5 compass. Location anchor for the candidate claims reaped from the 2026-06-06 neuroimmune-modulation / compulsive-loops thought intake. Off the V3 critical path. No substrate, no experiment in V3 until routed by an explicit version decision.
Seed
A saved Neuroscience News item — “Brain Immune Cells Drive Compulsive Behavior” (microglia-calcium-signaling-anxiety-ocd-30687; primary study Nagarajan et al., Molecular Psychiatry, 2026-04-13: Hoxb8 microglia intracellular Ca²⁺ signaling as a bidirectional optogenetic switch for anxiety/OCD-spectrum grooming) — prompted a single REE-relevant architectural idea:
Compulsive behaviour may be shaped by slow background substrate-state modulators, not only by fast cognitive beliefs, rewards, or action-selection loops. A loop can be hard to release because the field is biased, not because the goal is valued.
Precision guardrail (carried on every claim below). The paper’s mechanism is a fast / causal microglial Ca²⁺ switch, not a demonstrated slow tonic process. “Slow modulation” is the REE-side abstraction the saved item prompted, not the paper’s claim. Do not cite this study as evidence for slowness, do not reduce OCD/compulsion to inflammation, and do not imply REE or AI systems have immune cells — the analogue is slow substrate-state modulation, not biology.
What REE already owns (and why these are not duplicates)
| Layer | Existing REE claims |
|---|---|
| Compulsion mechanisms | SD-034 (closure operator / right-moment release), MECH-268 (dACC PE-saturation, urgency accumulator), SD-045 (action-chunk runaway-chunking → OCD ritual, Graybiel 2008), ARC-071 (policy-composition slow accumulator), MECH-124 (consolidation-loop → option contraction) |
| Slow neuromodulator layer | SD-037 (orexin-analog broadcast gain modulator — closest existing “slow control-plane modulator”), SD-036 (GABA decay), MECH-186/187/188 (5-HT goal-pipeline gain) |
| Slow physiological bias on the harm stream | SD-048 (sensitisation-spike = inflammatory-sensitisation analog; fatigue-drift = allostatic-load analog) |
| Decommit / commitment / residue | ARC-016 (commitment threshold), MECH-342 (maintenance-time decommit), IMPL-005 (residue persistence as structural memory) |
| Psychiatric failure-axis | MECH-088 (four-plane taxonomy; the OCD/compulsion axis) |
REE therefore already runs the mechanisms of compulsion and has a slow modulator slot. The genuinely-new content is (a) a value-INDEPENDENT route to loop stickiness — the existing slow modulators (SD-037) act on z_harm / drive-seeding and the existing PE-saturation (MECH-268) is fast and within-loop, neither carries explicit authority over the decommit-friction / release landscape; and (b) the observation that compulsion is best described as a composition over the scattered terms, not any single scalar.
Reaped claims (registered 2026-06-09, candidate / substrate_conditional / implementation_phase: v4 / version_relevance: v4_v5)
MECH-369 — Slow substrate-state modulator with authority over compulsive-loop stickiness
A slowly-varying, broadly-projecting modulatory state biases the decommit-friction / commitment-threshold / action-release landscape so loops become sticky or hard to release independently of represented goal value or fast prediction-error. Distinct from MECH-268 (fast, within-loop), SD-037 (gain over z_harm / drive-seeding, no release authority), and MECH-106 (value-driven threshold). Most naturally an amendment extending the SD-037 slow-modulator cluster with a decommit-friction authority channel. depends_on: SD-037, SD-036, MECH-268, SD-034, ARC-016, MECH-342, SD-048.
MECH-370 — Composed compulsion-risk readout
compulsion_risk = f(loop_reinforcement, threat_salience, residue_persistence, decommit_friction, slow_modulatory_state) — a unification of quantities REE already represents separately (loop_reinforcement ~ SD-045/ARC-071; threat_salience ~ harm stream/MECH-268; residue_persistence ~ IMPL-005; decommit_friction ~ ARC-016/MECH-342/SD-034; slow_modulatory_state ~ MECH-369/SD-037/SD-036). Asserts the composition is the right level of description for compulsion, letting governance ask “is this loop stuck because it is valued, or because the field is biased?”. depends_on: SD-045, ARC-071, MECH-268, IMPL-005, MECH-369, ARC-016, MECH-342.
Q-063 — Modulator-trapped loop vs valued repeated goal
How does REE distinguish a high-value, legitimately-repeated goal from a slow-modulator-trapped compulsive loop (release blocked by decommit-friction, not by value)? And can offline integration reduce stickiness without eroding the non-erasable harm residue protected by INV-004/INV-006? epistemic_category set substrate_conditional explicitly so narrow_open_question does not fire. depends_on: MECH-369, MECH-370, ARC-016, SD-034.
V3 vs V4 boundary
Containment-only in V3. The compulsion mechanisms (SD-034/MECH-268/SD-045/ARC-071) and the slow neuromodulator layer (SD-036/SD-037) are already implemented; the SD-033 OCD-axis is governed by sd033_governance_plan.md. Do not build a value-independent stickiness modulator or a unified compulsion-risk readout in V3 — a probe would be vacuous until the slow-modulator-class distinction is actually built. Promote/demote is suppressed by substrate_conditional; these do not pressure V3 governance.
Cross-repo (NOT claims.yaml)
The AI-cognitive-failure-taxonomy analogue — a slow-modulator axis beneath named failure modes (context-window pressure, memory weighting, tool-retry pressure, unresolved alerts, resource constraints, explicitly substrate-state and not biology) — belongs in Latent-Fields/ai-cognitive-failure-taxonomy, tracked there, not here.