ClaimLedger: Verifier-Governed Claim Maintenance for AI-Assisted Research

Large language models can generate plausible research claims faster than a human can check the claims, their dependencies, and the evidence cited for them. The resulting failure is not only factual hallucination: it is state corruption in a living research corpus — confidence is promoted without support, retired results reappear, and scripts that merely print a number acquire the status of tests. We present Verifier-Governed Claim Maintenance (VGCM), an architecture in which a generator may propose claim-state transitions but cannot apply a promotion to the authoritative record. We formalise a ledger L = (C, D, B, T, S, H) comprising stable claims, load-bearing dependency edges, evidence bindings, a corpus-defined confidence order, lifecycle state, and an event history. Six executable invariants govern tier support, retraction propagation, evidence admissibility, reference integrity, history/state consistency, and principal authority. The deterministic audit core is separated from evidence execution and from the human commit decision. We implement the architecture in ClaimLedger, a pure-standard-library Python package with corpus adapters and policy hooks. Eight seeded fault operators are all detected by their assigned verifier layers and a clean fixture produces no findings; these are mechanism-coverage results, not field estimates of precision or recall. A non-physics experiment-log adapter, a measured scaling study, and a parity audit against a maintained physics corpus identify both the reusable kernel and a load-bearing adapter requirement: dependency edges must encode support, not merely citation. The system enforces consistency, not truth. Its contribution is a reviewable claim-state machine that remains useful when the generator is unreliable.