Completion Is Not Declaration

AI-assisted research systems increasingly rely on generated status claims such as "complete," "reviewed," "checked," "verified," "ready," and "final." These claims are operationally useful, but they are not equivalent to verified research state. A system may correctly generate content while incorrectly classifying the state of that content, producing completion false positives and other status false positives that can propagate through research workflows. This paper introduces completion-claim governance as a distinct layer within transcript-sufficient and AI-assisted research systems. Building from a motivating case involving truncated transcript ingestion and incomplete source visibility, the paper develops a formal distinction between declaration, verification, acceptance, and enactment. It argues that status claims should be treated as governed research objects subject to audit rather than as self-validating declarations. The paper introduces a status-false-positive taxonomy, a state-transition ladder, and the DECLARED_COMPLETION_AUDIT operator as a bounded monitoring procedure for evaluating completion claims before they alter research state. It further examines the role of monitoring agents, multi-agent verification structures, and trust-but-verify governance architectures. The central claim is that completion claims are fallible outputs. A declaration does not establish completion, verification, acceptance, or enactment. Trust may route attention, but governance requires verification.