Executable Governance at the Point of Irreversibility A Formal Architectural Framework for Authority Enforcement in Composed AI Systems

As AI systems evolve from bounded models into agentic, multi-system architectures,governance failures increasingly occur despite local compliance of individual components.These failures are structural rather than behavioral, arising when legitimate authority cannot be demonstrated or enforced at the moment composite actions become irreversible.This paper introduces a formally specified architectural governance framework comprisingfour integrated mechanisms, ISDAIRE (ex-ante admissibility), ARETABA (runtime enforcement), GAG (compositional authority preservation), and MGAG (multi-layer governance).Authority is treated as a first-class executable object, defined ex-ante, cryptographicallybound at runtime, compositionally preserved across system boundaries, and deterministically enforced at irreversibility boundaries. The framework is mathematically rigorous, falsifiable, implementation-agnostic, and applicable to any distributed architecture producingirreversible outcomes. Formal definitions, atomicity semantics, provenance requirements,integrity-based audit criteria, and bounded latency analysis are provided. This work isproposed as a reference architecture for executable governance in composed AI systems.