Execution-Time Authorization for AI Agents: A Formal Framework for Deterministic Governance Boundaries

Execution-time authorization is a deterministic enforcement layer that evaluates canonicalized action instances against versioned policy and governed system state prior to execution, producing a replayable authorization verdict and an immutable decision artifact. This paper formalizes execution-time authorization as a distinct architectural category within AI governance—irreducible to guardrails, alignment techniques, identity and access management, observability tooling, or agent orchestration. As autonomous AI agents increasingly execute irreversible actions in regulated domains—including healthcare, finance, defense, and critical infrastructure—the gap between what an agent is technically capable of doing and what it is authorized to do under governing policy becomes a liability boundary. Execution-time authorization closes this gap by interposing a mandatory, fail-closed decision boundary at the point of irreversible state transition. This paper makes five contributions: (1) a canonical definition specifying the necessary and jointly defining properties of execution-time authorization; (2) a formal mathematical model of the authorization function with determinism, canonicalization, and replay invariants; (3) testable conformance criteria that distinguish conforming execution-time authorization from adjacent governance categories; (4) a structured architectural analysis differentiating authorization from guardrails, alignment, IAM, observability, and orchestration; and (5) integration and deployment patterns for implementing execution-time authorization within enterprise AI infrastructure. Execution-time authorization is established as a foundational primitive for enforcement-grade AI governance, providing pre-execution, deterministic, and independently verifiable decision control for autonomous systems.