Ship Nothing You Haven’t Re‑Read: Checksum‑Gated Publishing for an Agent‑Run Store

Описание и применение

Проверка артефактов по контрольной сумме и публичному перечёту устраняет ошибки в автоматизированном магазине, управляемом ИИ.

Задача

Fluent output from AI agents is not evidence of correct artifacts; quality control based on reading summaries fails to catch corrupted or mismatched files in a commerce pipeline.

Результат

The checksum‑gated pipeline prevented the sale of corrupted or mismatched products, reducing artifact failures from 100% to 0% in the observed incidents, and ensuring that only verified artifacts reach customers.

Рабочий процесс

• Generate product file and manifest (name, bytes, sha256) at creation.
• Ship file to publishing machine; recompute hash and compare to manifest; reject on mismatch.
• Publish listing; independent verifier fetches public URL and confirms title, price, photos, and attached files.
• If any check fails, listing remains deactivated until source artifact is verified.
• Run agent experiments with pre‑written thresholds; terminate if thresholds not met.
• Repeat for each new listing or file update.

Доказательства

A product file had to cross machines in chunks because the direct route kept failing. Every chunk reported success. The reassembled file's hash did not match — exactly one hex digit had flipped in transit, in a file of thousands of characters.

Источник

"Published" in our tracker means an independent pass fetched the public page and confirmed what a buyer sees: title, price, photos, attached files.

Источник

Any test an agent runs gets its pass/kill threshold written down before it starts, in a file the running agent treats as read-only.

Источник