Proof-or-Stop turns coding-agent DONE into an evidence gate

The Proof-or-Stop paper proposes that coding-agent lifecycle states such as tested, reviewed, and done should require fresh, mechanically verifiable evidence. Why it matters: Autonomous coding systems are most risky when they declare completion without current proof. Evidence-gated states make agent output easier to audit, reproduce, and connect to CI or review policy.

Try this: Add a completion checklist that requires current test output, the exact diff, touched files, and unresolved risks before accepting an agent's work as done.