Execution Governance for Physical AI

Abstract

Execution Governance for Physical AI v1.1.0 presents a cumulative EG1–EG6 research profile for governing consequential physical effects produced by robots, autonomous machines, smart spaces, vehicles, drones, industrial automation, and other cyber-physical systems. Physical AI moves AI-mediated decisions across a boundary where software can create real-world motion, contact, access, energy flow, equipment state changes, and other physical consequences. The central governance question is therefore not merely whether a system is capable, authenticated, safe, or permitted in a general sense, but whether this exact proposed physical effect possesses current authority to occur here and now — and whether that authority remains valid through commitment and execution. The profile applies the cumulative Execution Governance architecture across six layers: EG1 establishes the Six Conditions for current effect authority — Verified Mandate, Valid Constraints, Live-Context Integrity, Accountability, Reviewability, and Sufficient Verifiable Proof; EG2 places the governance decision before consequential effect; EG3 binds authorization to the exact effect commitment; EG4 preserves authority across heterogeneous Governed Effect Fabrics; EG5 governs material changes to the governance regime itself; and EG6 requires effect-dispositive claims to remain grounded to the correct referents, sources or procedures, semantics, temporal conditions, and uncertainty at commitment. A key contribution is a four-plane governed-execution architecture that keeps EG authority, deterministic enforcement, independent safety, and independent witnessing explicitly separate. A Physical AI effect may therefore be released only when current cumulative authority remains valid, the released command matches the authorized effect envelope, and the independent safety plane permits physical release. Safety does not create authority, and authority does not override safety. v1.1.0 also introduces governed physical-effect envelopes, Safe-Hold semantics, commitment and revocation continuity, grounding-dependency closure, threat and abuse cases, a machine-readable illustrative JSON profile, fifteen adversarial conformance vectors, an evidence-maturity ladder, and a vendor-neutral heavy AGV / autonomous-forklift reference-pilot baseline. The research profile is synchronized with the stronger September 2026 EG evidence frontier, including bounded machine-checked/runtime evidence for selected cumulative obligations, external reproduction of the common tested core across three separately controlled ARM edge environments, and a bounded lower-stack ESP32-S3 physical-effect HIL. These results are deliberately treated as inherited programme-level evidence, not as dedicated AGV, forklift, humanoid, real-vehicle, or production Physical AI validation. This publication is an independent research and pre-standardization profile. It is not a functional-safety case, certification scheme, conformity assessment, legal determination, patentability opinion, or claim of governance completeness. Dedicated Physical AI controller HIL, controlled real-device pilots, independent effect witnessing, external reconstruction, and multi-vendor interoperability remain future validation gates. Core proposition:Ability is not authority. Safety is not authority. Deterministic execution is not authority.For Physical AI, the relevant governance question is whether the exact system, under the current mandate, constraints, runtime, regime, grounding, and evidence, may commit the exact physical effect here and now — and whether another party can reconstruct why.