AI Can Write the Code. Can It Prove the Fix?

Описание и применение

Автономные агенты могут генерировать «зеленые» патчи, которые на самом деле неверны. Проект no_human вводит строгие проверки: защита от изменения тестов, воспроизводимость бага, независимый ревьюер и правила слияния, чтобы убедиться, что изменения действительно исправляют проблему, прежде чем человек их примет.

Задача

Autonomous coding agents can produce green diffs that are actually wrong, leading to costly review cycles and undetected bugs.

Результат

By separating implementation from review, enforcing deterministic gates, and requiring human merge approval, the workflow reduces the risk of deploying incorrect AI‑generated code and improves trust in autonomous coding.

Рабочий процесс

• 1. Submit a ticket to the system.
• 2. Agent generates a plan and implements changes locally.
• 3. Agent runs tests and adjusts them if necessary.
• 4. Tamper guard checks for test count drops, skip markers, tautological assertions, and autouse fixtures.
• 5. Reproduction gate verifies that the bug test fails on the base commit and passes after the change.
• 6. Independent adversarial reviewer opens a fresh session, refuses edit tools, and produces a checklist of findings and a pass/fail verdict.
• 7. Held‑out tests and verifiers run as evidence unseen by the implementer.
• 8. Merge policy evaluates all gates and sets a ready flag.
• 9. If ready, a PR is opened and parked at awaiting_approval; a human must merge.
• 10. If any gate fails, the loop stops, escalates, or retries up to limits.

Доказательства

The project’s own documented pipeline: ticket → context → plan → implement → review → test → PR → you merge

Источник

Tamper guard diffs test files separately from product code, and fails on a net drop in test count or assertion count

Источник

Reproduction gate takes the tests the coder offers as evidence, copies them into a worktree at the merge base, and requires them to fail there and pass on the new tree

Источник