Persistent AI agents do not run as stateless functions; they revise beliefs, consolidate memory, and swap out the substrates they execute on. Each of these changes produces a successor state. The problem, according to a new arXiv paper, is that similar successor states can accompany differently authorized transition claims. A state that looks like a legitimate development may in fact be the result of an unauthorized change, and a legitimate change may alter the state substantially.
To address this, the authors propose the Cognitive Continuity Test, a method for verifying governed state transitions. The test is meant to check whether a transition from one agent state to another is authorized and governed, rather than merely similar in surface features. The paper's framing suggests that continuity cannot be established by output similarity alone; the governance of the transition itself must be examined.