A vendor “double-check your work” prompt asks the model to re-read its own reasoning; this one attacks the specific failure where an agent says “tests pass” after only reading the assertions. It grades the claim by how it was actually obtained and refuses the sign-off words unless a proof line – tied to evidence visible in the same conversation – comes first, born from a real case where a confident “tests pass” summary reached a pull request before anyone noticed nothing had run. It is the sign-off twin of the claim-ladder gate: claim-ladder governs what you encode into code and config; this governs what you tell a human when you close the task.
Fill in: sign_off_claim
Shared as is, for reference. Read it and decide what it will do before you run it; using it is your responsibility, under the terms.