As AI systems evolve from bounded models into agentic, multi-system architectures, governance failures increasingly occur despite local compliance of individual components. Thesefailures are structural rather than behavioural, arising when legitimate authority cannot bedemonstrated or enforced at the earliest governed commit boundary that binds a composedaction to an irreversible outcome. This paper introduces a formally specified architecturalgovernance framework comprising four 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, cryptographically bound at runtime, compositionally preserved across multidomain system boundaries, and deterministically enforced. Crucially, where actions exposemultiple candidate points of irreversibility across domains, the architecture dynamicallybinds enforcement to the uniquely identified earliest governed irreversible execution boundary.The framework is formally specified, falsifiable, implementation-agnostic, and applicable todistributed architectures producing irreversible outcomes where the earliest governed commitboundary on the system side is interceptable and can be atomically coupled to enforcement.Formal definitions, atomicity semantics, provenance requirements, integrity-based audit criteria, and bounded latency analysis are provided. This work provides a reference architecturefor executable governance in composed AI systems. Its guarantees are enforcement-correctrelative to declared governance artefacts and declared irreversibility surfaces; it does notinfer omitted boundaries, omitted dependencies, or incomplete authority predicates, andtherefore is not complete with respect to reality.
Masayuki Otani (Sat,) studied this question.