This paper defines the authorization threshold for governance in AI systems. It argues that governance is not a property of systems in general, but of specific decisions made under specific constraints. As AI systems increasingly initiate actions autonomously, operate across institutional and regulatory domains, and produce irreversible outcomes, governance must be evaluated at the level of action validity rather than system behavior. The paper distinguishes procedural approaches to AI governance—monitoring, alignment, filtering, logging, and observability—from the definitional conditions required for an action to qualify as governed. It proposes that an action is governed only if its authorization can be established prior to execution under explicit, evaluable constraints and made available for independent verification. To formalize this standard, the paper introduces the Canonical Action Frame, defines the distinction between interpretive systems and authorization systems, and draws a sharp line between evidence and proof. It argues that most governance architectures omit the critical transformation layer between policy and evaluation: the constraint representation layer, where human-readable rules are formalized into canonical, machine-evaluable constraints. The paper further defines seven conditions of governance validity and five dimensions of governance completeness, including constraint completeness, semantic precision, policy binding, temporal binding, and independent replayability. It concludes by specifying the structure of a valid authorization artifact and proposing a testable standard for determining whether a decision was genuinely governed. This work is intended for researchers, regulators, enterprise buyers, compliance leaders, technical architects, and others evaluating what it means for AI governance to be formal, enforceable, and provable rather than merely observable or descriptive.
Edward Meyman (Sat,) studied this question.