The Zombie Intent Hazard: A Safety Violation in Write-Ahead Audit Architectures for Safety-Critical Systems

This publication identifies and formally solves a previously undocumented safety vulnerability in write-ahead audit architectures: the Zombie Intent Hazard. This hazard occurs when a system persists an intent record to non-volatile memory, crashes before actuation, and upon reboot executes the stale intent despite changed environmental conditions. We demonstrate that this vulnerability was missed by three frontier AI systems (GPT-5.2, Claude Opus 4.5, Gemini 3) during independent architectural review. We present a formal solution using session-bound verification and prove its correctness using TLA+ temporal logic.