Two Structural Boundaries of Governance in Autonomous Systems — Paper 9 This paper formalizes a structural view of AI governance based on two distinct boundaries: the admissible state-space boundary and the execution (commit) boundary. Most governance approaches operate post hoc, evaluating or constraining decisions after they are generated. This work argues that such approaches are structurally insufficient, as they act over a proposal space that remains unconstrained. We introduce a dual-boundary model: Admissible State-Space Boundary: defines what can exist in the system at all. Governance at this level is not enforcement but construction — certain states are never representable. Commit Boundary: governs the point of irreversibility, where actions become real and must be explicitly authorized, non-bypassable, and evaluated at execution time. The paper shows that: Governance applied only at execution is reactive. Governance applied only at the state-space level lacks enforcement guarantees. Robust governance requires both boundaries, structurally coupled but functionally distinct. This work reframes governance from a control layer into a property of system architecture, shifting from filtering behavior to defining possibility.
Ricardo Rubio Albacete (Thu,) studied this question.