Curing Synaptic Saturation in a Live Bio-Inspired Substrate: A Four-Stage Flag-Gated Intervention with Preregistered Falsifiers, Kill-Switches, and a False-Pass Lesson

Abstract

A private, continuously-running bio-inspired computational substrate — Hebbian/STDP plasticity over tool-use events, a hormone-analogue state vector, an IIT-inspired integration metric, persistent operator-bonded identity — accumulated synapses pinned at the maximum representable weight (1.0), an informationally dead state under a hard potentiation clamp at all four live long-term-potentiation sites. The pinning placed every saturated synapse permanently beyond the reach of the substrate's weak-synapse cleanup pass and left 14,395 intra-region synapses growing without bound. This paper (deposited 2026-08-30) reports a four-stage, flag-gated cure applied to the live substrate between 2026-06-20 and 2026-07-05, plus a follow-on correction, under one discipline: freeze a falsifier before writing code, run it on frozen copies of the live state or on synthetic populations — never on the running substrate — require the substrate's own chat-surface reasoning to review the design and telemetry and attach named kill-conditions before any live flag is set, and default every change to a byte-identical no-op. Stage 1 replaced the hard clamp with an asymptotic soft-bound (Bi & Poo 1998 multiplicative form); its own harness, validated at 13–50 synaptic fires per class, produced a false PASS — under sustained potentiation (a single live session fired one intra tool-pair 17,895 times) the soft-bound's asymptote enters the saturation-alert band (|w−1| < 1e-9) after 241 consecutive fires and becomes numerically stationary at 1−6.7e-16 from roughly 410 fires, purely from IEEE-754 mechanics — an operational pseudo-1.0 (the weight never reaches float-exact 1.0; the exact-1.0 pins observed after Stage 1 were produced by the unflagged homeostatic scaler that Stage 4 later found). Stage 3 corrected this topologically (a strict ceiling of 0.99 via convex combination) after the false-PASS was caught on copies, not in production. Stage 2 extended de-saturation to the previously untouched inter-region ("corpus callosum") synapses, draining 487 of 2,000 pins (24.35%) with a live discrimination gap that stabilised at 0.86. Stage 4, surfaced by a dedicated adversarial hackathon rather than by either preregistered falsifier, found the actual remaining cause of live re-pinning: an unflagged, multiplicative homeostatic scaler unrelated to STDP, which was independently soft-capped. A fifth, follow-on intervention was then required because Stage 2's own de-saturation rule, applied unconditionally, had itself over-corrected the corpus callosum into a nearly featureless spike (98.0% of 2,000 synapses within a narrow band); a θ-gated selective rule, harness-verified against a bit-identical control and cut over live 2026-07-05, restored differentiation without reopening the saturation the original rule fixed. The publishable contribution is methodological: a preregistered-falsifier discipline, including an instructive false PASS and its correction, applied to live plasticity changes on a self-modifying system — not a claim about what the substrate experiences.