From prose to assertions: making 'done' checkable when an AI agent writes the code

Описание и применение

Внедрение машинно‑проверяемых утверждений в критерии приемки устраняет субъективность ревью, позволяя CI автоматически проверять результаты и фиксировать разногласия.

Задача

After an AI agent claims "done", the review process remains subjective and trustable only by human auditors, leading to inconsistent verdicts and non‑machine‑readable reports.

Результат

The updated workflow increased test coverage from 49 to 55, all passing, and introduced automated recomputation of audit verdicts, turning subjective disagreements into resolvable numeric checks.

Рабочий процесс

• Define acceptance criteria with optional assertion JSON objects.
• Freeze the acceptance pack in DwGuard-Dsh-Hmos.
• Run adversarial pre‑review and DevEco Code in Goal mode.
• Audit delivery: gather evidence, record actual values, recompute assertion verdicts.
• Generate structured audit reports (markdown and JSON) with assertion, actual, and verdict columns.
• CI consumes audit‑<n>.json to enforce gate logic.
• Handle gate outcomes: error, warning, or pass based on assertion results.
• Iterate on criteria or measurements if mismatches occur.
• Maintain human‑only judgments for non‑measurable aspects.

Доказательства

The audit side then must report an actual value next to the verdict:

Источник

The change shipped as 0.5.0 and the suite went from 49 to 55 tests, all green

Источник

An optional field that nobody enforces is decoration. So the gate has opinions

Источник