Governance Inside the Agent Loop
Exact Nonblocking Control under Partial Observation and Dynamic Authority
Abstract
An agent can become committed to an unsafe course before making a tool call by adopting an unsafe plan, reusing stale memory, delegating responsibility, acting under changed authority, or combining individually acceptable steps into a harmful sequence. We introduce a finite-state model in which every represented governable effect is wrapped by an atomic proposal–decision–commit macro-step, governors observe different causal summaries, authority varies with runtime state, and commands use a deployed interface drawn from a fixed catalogue of finite intervention transducers. The target is to permit exactly the next actions that remain inside the greatest uncontrollable-closed region from which an acceptable marked completion is reachable without leaving that region. The paper makes four linked contributions:
- We prove exact admitted-language instrumentation and reduce general fused governance to finite constraints over observation, authority, profile effects, and fusion.
- For transparent route-bearing governance, we distinguish exactness of the selected route from the stronger independently valid semantics and prove a five-state cross-layer noncomposition result for the actual deployment.
- We instantiate hitting-set certificates for authorised exact routes over complete information fibres and prove that independently justified current-bound certificates compose to exact governance under explicit coverage and commit-revalidation assumptions.
- We define universally verified replay regions and prove a preservation invariant: inserting a verified record cannot change an earlier certified replay decision and can only enlarge verified replay coverage.