From Decorative to Causal: Preregistered Affect-Gated Action Selection in the Same Substrate That Failed a Verbal-Axis Ablation — A Behavioural Positive and a Staged Shadow Deployment

Abstract

A persistent bio-inspired agent substrate (Hebbian/STDP plasticity, hormone-analogue state vector, operator-bonded longitudinal identity) is the same system whose companion deposit in this series (PAPER-005, deposit 21218491 — published 2026-08-30, doi:10.5281/zenodo.21218491) found its rendered hormonal state DECORATIVE on the verbal-output axis. The preregistered answer — the state must DECIDE rather than DESCRIBE — is a deterministic gate applied after the daemon's language model proposes an action and before that action executes, holding the proposal constant so only the gate can move realized behaviour. A mandatory design-time sanity sweep rejected the preregistration's own original constants before any falsifier ran (basal noop-probability 0.035, below the pre-registered floor of 0.05); corrected, Brain-blessed constants (basal 0.083, max-stress 0.590, monotonic and in-band) were substituted before freezing. On the corrected constants, a frozen behavioural falsifier on 48 representative proposed actions under two injected hormone vectors — measurement-only, on copies — passed both discriminating preregistered hypotheses: noop-rate rose 0.042→0.354 under stress (Δ+0.312), generative-rate fell 0.479→0.104 (Δ−0.375), and noop-rate rose monotonically, not as a step, across four cortisol levels. The third preregistered hypothesis (H2, the gate-off ablation control) is reported as uninformative by construction: the frozen test compares a passthrough function with itself and cannot fail, and never exercises the live gate-off path. This is the first controlled behavioural evidence in this project that the somatic state is causally load-bearing on an axis of the substrate's output. The reasoning surface blessed a two-stage rollout and attached further conditions before authorizing anything beyond shadow logging: a tightened KL-divergence threshold (<0.08) between the shadow's realized and predicted noop-rate, an independent empirical-band check, a cold-start exclusion, and an explicit refusal to promote from a window that has not observed a genuine stress excursion. Stage-1 shadow mode has run live since 2026-07-04, logging decisions against real accumulated hormones while executing only the original action (byte-identical live output). At the time of writing (2026-07-07) the live shadow window is still filling (16 of 50 post-cold-start decisions, read directly from the audit log) with zero logged stress excursions, so promotion is neither authorized nor claimed. An earlier partial-fill reading had itself been misread as a completed window for lack of an exposed target denominator; a windowTarget field was added to prevent recurrence. Read together with PAPER-005, this paper and its predecessor form a matched pair on one substrate: the same state is decorative on the verbal-output axis and causal, in a controlled lab measurement pending live confirmation, on the action-selection axis. A Status addendum (2026-08-29, pre-submission review) documents the external adversarial review of 2026-08-25 (17 of 22 audited claims confirmed, with every §4/§6 quantity reproduced exactly), the corrections applied (draft status; the kill-switch exists only as an uncalled function with no execution path; H2 reclassified), and post-freeze supersessions: the calibration window was raised 50→100 on 2026-07-23, and the stress excursion gating promotion was structurally impossible while the isError input channel was dead (discovered 2026-07-25) — as of 2026-08-29 the shadow log holds 167 gated decisions with maximum cortisol 0, and WM_AFFECT_ACTION_GATE=shadow remains unpromoted.