Policy-as-Architecture: Treating AI Safety as a Kernel-Level Resource Management Problem

Current AI safety practice relies almost exclusively on application-layer, linguistic guardrails that a sufficiently motivated adversary can bypass through adversarial formatting or novel prompting. This creates what the author terms "Accountability Evaporation": a structural responsibility gap in high-stakes deployments where no deterministic, hardware-enforced boundary exists between an AI agent's proposed action and its execution in the physical or consequential layer. The Autonomous Safety Kernel (ASK) is a three-layer architecture that decouples non-deterministic AI proposal generation from deterministic, hardware-gated execution: Layer 1 produces canonical, cryptographically signed action envelopes; Layer 2 evaluates them deterministically against governance-signed rule bundles; Layer 3, implemented as an Enforcement Microcontroller with hardware root of trust and independent power domain, controls the execution bus and cannot be bypassed by the Agent Host. The Global AI Safety Index (GASI) proposes a consortium-governed benchmarking process model — modeled on the SPECpower precedent — defining a Safety-per-Outcome metric class and a graduated adversarial load protocol with anti-gaming controls. Three concrete embodiments are provided: operational technology control systems, financial transaction gateways, and healthcare patient safety. A provisional patent application has been filed (U.S. priority date: June 30, 2026).