Architectural Constraints in AI Governance: Execution Legitimacy, System Trajectories, and Structural Boundaries in Probabilistic Decision Systems

This paper addresses the governance of AI-enabled operational systems from an architectural perspective. It argues that effective governance cannot be achieved through behavioural supervision or post-hoc auditing alone, but must be embedded directly within the system architecture that determines which state transitions are admissible at the moment of execution.The paper introduces the concept of trajectory-level governance, which evaluates system legitimacy not only at discrete decision points but across evolving sequences of transitions. Drawing on safety engineering literature and systems-theoretic approaches, it identifies a class of governance failure modes — including interpretative governance, epistemic drift, compliance illusion, and trajectory divergence — that arise when governance mechanisms remain external to the execution architecture.The paper further introduces the Constitutional Failure Boundary (CFB) stress-testing methodology, applied to reconstruction scenarios based on historical system failures including the Knight Capital trading incident (2012) and the Boeing 737 MAX MCAS control system failure. Results suggest that execution-constrained architectures can prevent recursive amplification behaviour under degraded legitimacy conditions.The paper concludes with a formal representation of execution admissibility and a set of architectural principles for designing governance-capable AI systems across financial, industrial, and healthcare domains.