Ethical Hyper-Velocity (EHV): A Hardware-Rooted Zero-Trust Runtime Enforcement Architecture for Agentic AI Systems 文章

arXiv:2605.17909v2 Announce Type: replace Abstract: As autonomous agentic systems scale across regulated critical infrastructures, the lack of mechanistic, hardware-rooted enforcement for high-frequency policy updates presents a fundamental safety gap. We present Ethical Hyper-Velocity (EHV), a governance-aware runtime enforcement architecture for agentic systems that combines Grammar-Constrained Decoding (GCD) for inline policy-constrained token generation, Causal Graph CRDT-based policy synchronization with vector-clock ordering, hardware-attested execution in Trusted Execution Environments (TEEs), and OSCAL-formatted machine-readable audit logging. Unlike retrospective auditing frameworks (ISO/IEC 42001, NIST AI RMF) that introduce 14-30 day policy latencies, EHV relocates the Policy Enforcement Point (PEP) into the inference pipeline via a Governance-Aware Just-In-Time (JIT) Compiler.