Turn "done, fixed, tested" into evidence a second agent judges. Use when you are about to report a task as complete, when an agent or teammate claims something works, or before acting on a "tests pass / 0 findings / all green" report. The claimer captures raw command output with a re-runnable command (scripts/evidence.py); an auditor who did not do the work reads only the evidence and returns one of three verdicts — supported, not supported, insufficient — never a fourth. Includes the table of things that look like evidence but are not.
