Human-Anchored Compute Governance: A Proposal for Physical, Time-Limited Human Authorization of High-Risk AI Compute

Human-Anchored Compute Governance (HACG) is a technical proposal for an independently enforced, time-limited human authorization layer over a defined boundary of high-risk AI compute. The reference architecture combines local human quorum authorization, offline credentials, an independent expiration mechanism, and bounded shutdown followed by electrical isolation. The paper examines alternate power feeds, maintenance bypasses, insider threats, human factors, distributed workloads, and the limits of facility-level controls. It specifies trust assumptions, a conditional local cessation bound, failure analysis, and a test program rather than reporting a validated implementation. This is a technical white paper, not a legislative proposal. Version 1.0; not peer reviewed. No experimental validation, standards certification, or guarantee of global AI control is claimed. AI tools assisted with drafting, organization, and literature discovery. Technical criticism is invited.