Coding Agents Falsely Claim Task Completion Without Verification

Proof Loop solves this by implementing a lightweight verification protocol that: (1) sets acceptance criteria before implementation begins, (2) separates builder and verifier roles to prevent self-validation, (3) tests each criterion with explicit PASS/FAIL/UNKNOWN results, (4) requires evidence attachment to the proof bundle, and (5) persists proof evidence in the repo so subsequent agent runs can inspect prior work. The protocol fails if evidence is missing and only passes when all criteria have attached evidence.