Verified Semantic Transitions (VST): Belief-Gated State Change and Repair-Invariant Memory for Reliable Long-Horizon AI Agents

Tool-using AI agents increasingly operate in asynchronous, partially observable environments where a successful API return, mouse click, keystroke sequence, or tool invocation does not guarantee that the intended semantic outcome occurred. This white paper proposes Verified Semantic Transitions (VST), an implementation-neutral architectural primitive that treats the unit of reliable agent progress as a verified intended state change rather than an attempted action. A VST binds an intent, prior belief state, predicted postcondition, action, resulting observations, verification predicate, recovery policy, and memory update. The central rule is that external action completion must not automatically become an internal fact: an agent should advance its plan or promote a proposition into working belief only after sufficient evidence establishes semantic closure. The paper formalizes this rule, introduces a three-clock model of action, environment, and cognition timing, defines a taxonomy of transition failures, and proposes repair-invariant memory as a mechanism for storing reusable corrections rather than only raw trajectories. VST is not presented as the invention of transactions, feedback control, Hoare-style postconditions, belief-state estimation, or agent verification. Its contribution is integrative and architectural: it unifies these ideas at the transition level for stochastic agent runtimes and couples verification failure to reusable repair abstraction. The framework is positioned relative to ReAct, Reflexion, Voyager, ToolGate, Estimate-Verify-Update (EVU), SAVeR, Agent-BRACE, Lean4Agent, and Cordon. In particular, VST is complementary to task-level semantic transaction systems: task transactions govern durable effects, whereas VST governs whether a specific semantic change may be committed into agent belief and control flow. The paper concludes with measurable metrics, a proposed ablation protocol, and an implementation-neutral schema suitable for desktop, browser, API, robotic, and enterprise agents.