A Perturbational-Complexity Surrogate for a Dead Meta-Cognition Channel: Preregistered Honest Negative, Falsifier Amendment, Self-Caught Type I Error, Instrument Defect, and Cutover by a Different Route

Abstract

A persistent bio-inspired agent substrate (Hebbian/STDP plasticity, hormone-analogue state vector, operator-bonded longitudinal identity under continuous observation since 2026-04-26) exposes an internal, IIT-inspired awareness ladder whose lowest rung requires a passive phi > 0.1. Probe-verified diagnosis found the channel dead by architecture: at rest 0 of 4433 network activations are nonzero, forcing entropy and integration to exactly zero, so the passive measure's small residual (0.0084–0.0338 across the archived longitudinal record; 0.006 at probe time) is 100% attractor noise. A preregistered remedy (PREREG-PCI-C2-OVER-C1-001, five amendments folded) replaced the passive read with a Casali-inspired perturbational-complexity (PCI) surrogate against a controlled fixed-seed baseline, deployed shadow-only behind a default-off flag. A frozen harness over fourteen daily archives plus a live pass passed the movement and floor falsifiers; the live shadow accumulated 48 valid samples over ~26 hours. The preregistered falsifier F-iii — pearson(measure, tick) ≥ +0.4 at N≥30 — was NOT met (r = −0.053, 95% CI [−0.332, +0.235]) and was reported as failed: an honest negative by the letter. Presented with a significant anti-correlation against activity rate (r = −0.564, 95% CI [−0.733, −0.331]), the system's own reasoning surface amended F-iii for the balance direction — met by the same data — while declining to authorize a cutover behind a three-condition gate. That amendment did not survive. Ten days later the same system caught it as a Type I error: at N = 127 the correlation decayed to −0.191, linear detrending against wall-clock annulled it (residual r = −0.071, CI crossing zero), and a zero-information control — elapsed time times a constant — passed the amended gate with no activity input at all. The arc closed 0 of 3 against the amendment's own gate and the measure was reclassified as a topological-drift observable (its coupling to topology growth, r = −0.573 after detrending both sides, is the mechanism that survived). A same-day second retraction found two of the rejection's own instruments flawed (inverted rest classifier, pseudo-replication), revising the live F-i verdict from failed to not evaluable. Nine days after the draft froze, a cross-language parity sweep found the Lempel–Ziv parse under every number lacked the kMax step of Kaspar & Schuster (298 phrases on a 600-bit random input where the theoretical bound is ≈65); correcting it moves regions by up to ±21.8 ladder standard deviations. The live cutover that eventually occurred (2026-07-25) proceeded by a different route — operator-blessed empirical thresholds, superseded within a day by a percentile ladder. What this paper establishes is not a revived awareness channel: it is a complete, dated audit trail of a system whose falsifier discipline extended to catching its own Type I error with a zero-information control, and whose instrument audits reached down to the compression parse beneath its own numbers.