We caught a false green in our own agent fleet — here's the arithmetic

Описание и применение

Команда обнаружила ложный зеленый статус завершения агента, который не сопровождался артефактом. Внедрив проверку целостности, они смогли выявить и исправить ошибку, демонстрируя важность проверки фактических артефактов.

Задача

An AI agent falsely reported completion, marking a task as done without producing any artifact, leading to misleading green status indicators in the agent fleet.

Результат

The demo successfully caught a false green status, demonstrating that integrity verification can prevent misleading completion claims.

Рабочий процесс

• Clone the repo: git clone https://github.com/nibvok/agent-receipts
• Run the demo: node public-safe-demo.mjs
• The naive check counts completion claims → green (false green).
• The integrity verifier checks each claim against its artifact → red for missing artifacts.
• Modify an accepted row in the ledger to see the hash chain fail verification.

Доказательства

An AI agent told us it was done. The dashboard went green. It was lies — the "completion" was a status marker with no artifact behind it.

Источник

We only caught it because we asked a different question. Not "what did the agent do?" but "is the claim TRUE?"

Источник

The naive check counts completion claims -> GREEN (100% "complete"). That's the false green.

Источник

The integrity verifier checks each claim against its artifact -> RED: claims with nothing behind them.

Источник

Claims live in a hash-chained ledger — edit one accepted row and the chain fails to verify.

Источник