Why it matters
Fieldguide walks through a hypothetical bank reconciliation to show why a correct-looking answer is insufficient audit evidence. Controls prevent agents from overwriting human-reviewed work, record individual execution steps, bind citations to specific document versions and hashes, and check whether evaluation labels remain valid. The worked examples illustrate a control design rather than a measured deployment success rate or a guarantee of regulatory compliance.
My takeaway: Make changes to approved records new proposals requiring review. Store step inputs, outputs and stable evidence identifiers; compare changed steps when runs disagree. Revalidate historical evaluation labels after policy changes, and escalate conflicting evidence with enough context for a reviewer to investigate.