AI agents author implementation at rates that outstrip human review, and governance approaches inherited from human-authored software target the wrong artifact. We propose a different register: the governed artifact should be a formal contract over the system's declared intent, its permitted effects, and the human review points through which canonical decisions are made — not the implementation itself, which may take forms humans cannot efficiently read. We present Warrant, a governance contract model for AI-authored systems. Warrant has three distinctive elements: - A claim-first register expresses governance as machine-checkable propositions the system must uphold.- Canonical playbacks with no-silent-regression turn approved review artifacts into signed, hash-referenceable governance anchors that cannot be silently invalidated by later work.- Implementation opacity as a baseline provision states that representation is not a governance dependency; the framework accepts any representation the runtime can evaluate. We describe AgentFlow, a reference runtime that instantiates the contract against prose generation and application scaffolding, and report a record of AgentFlow governing its own development across multiple domains — an operational claim about tractability distinct from the empirical claim about efficacy that is the subject of a companion systems paper in preparation. We situate Warrant relative to ArbiterOS, EviBound, AGENTSAFE, SSGM, and D3, and argue that the integration of these elements is distinct from the prior work we examined.
Rod Bennett (Sat,) studied this question.