Margen

Margen tests corrections against what an agent actually stores. A caller changes an approved Wednesday booking to Thursday while a Muse Spark request is pending. The late response still proposes Wednesday. The host rejects the stale write, requires current approval and saves the current booking. The viewer separates the request, proposal, permission, stored effect and final reply. The developer contribution is executable: a three-operation host adapter, seven ordered probes, two unmodified implementations meeting all seven contracts, and a deliberately defective ledger that fails an ABA test. Wednesday changes to Thursday and back; the final day looks correct, but the consumed quote belongs to an old revision. The evaluator exposes the quote/effect mismatch. A separate 49-evaluation local mutation suite detects five deliberately planted fault types and includes a deny-all control that is safe but does not complete authorized work. The public reviewer includes the recorded Meta evidence, source-event replay, inspectable witnesses, exports and a reproducible offline verifier. Calendar, caller inputs and approvals are synthetic. Local diagnostic tests use programmed responses. This is not a live phone evaluation or a claim of production security, a new algorithm or competitive superiority. Earlier appointment-automation work inspired the problem; its original codebase and private systems are excluded.