prove-it.

Verify a completion claim against observed behavior. The verifier runs checks and records observations; the parent judges the claim. Verification alone authorizes no product edits.

Dispatch a capable verifier before choosing checks, honoring explicit user delegation and model choices; report a blocker if required delegation is unavailable. Run directly only for one prescribed cheap check or when none can be dispatched, and state why. Give the verifier the claim unchanged, the supplied facts and access limits, and the rules below.

Verify

Make the claim falsifiable: starting conditions, action or input, and expected observable result. Expectations come from the user or an independent contract, never from the implementation or its tests. A reported failure is reproduction evidence.

Run the smallest meaningful check at the boundary that owns the claim: public API, command, interaction, or stored result. For a fix, reproduce the failure first where practical, then run the project checks the requested scope requires. Reuse valid evidence until a change or gap makes it stale; keep failures visible.

Confirm the version, instance, and cases ran, and observe results and side effects. Mocks cover only mocked cases; builds and static checks leave runtime behavior unverified. Report checks and commands inline; save only when required.

Judge

A claim is verified when observations match, required checks pass, and failures are resolved, within the observed environment and cases. Failed, blocked, and untested claims stay unverified; send gaps back as bounded follow-up.

Present what ran, expected versus observed behavior, the verdict, and remaining gaps. Stop when the claim is verified or its failure or blocker is established.