UAS-OS: A Transactional Governance Kernel for Verifiable Autonomous Execution

Autonomous AI agents and intelligent systems increasinglyoperate in environments requiring dynamic, high-stakes decisionmaking.Traditional security models—such as static Role-Based AccessControl (RBAC) and conventional attribute-based policies—donot by themselves explicitly model execution-time causal state, dynamicallychanging intent, or post-authorization execution evidence.In this paper, we present UAS-OS, a kernel-level transactional governanceabstraction that unifies dynamic intent-driven authority, multi-FSM governance state machines, execution witness chains, and offlineexecutable invariant-based verification. We introduce a formal statetransitionmodel where authority is dynamically computed as a resourcespecificvector based on intent, evaluated risk, and trust telemetry.Furthermore, we formalize execution causality through strict witnesschains and resolve temporal ambiguities via three execution race safetyproperties. To address audit failures without destructive process termination,we introduce the Commit-In-Doubt state, unique recovery transactionmapping, and an explicit recovery protocol enforcing attemptlocalsemantics. Our invariant verifier (I1 ∼ I6) uses audit logs asreplayable execution traces under the implemented transition semantics,enforcing validate-before-mutate checks and strict per-attempt transactionserialization to enable end-to-end offline executable verificationof the modeled governance properties. Deterministic fault-injection andquantitative evaluations show that UAS-OS provides invariant-checkedexecution governance under the evaluated fault model, with a measuredend-to-end runtime overhead of 7.6%.