Strategic Audit Cards: Incentive-Compatible Disclosure for AI Evaluation Reports

Current AI documentation artifacts (e.g., model cards) are largely self-reported by developers optimizing for minimal liability. We model this as a "Strategic Audit Card" game: a signal extraction problem with partial verifiability. We demonstrate that mandatory minimumfield policies can drive "Compliance Theater," where developers fill templates with verifiable low-signal content while omitting hard-to-verify risks. We show uniform random audits are inefficient against strategic omission, whereas our proposed "suspiciousness-based" audit, targeting high-positivity, low-verifiability claims, improves truth-telling. Reliable transparency requires coupling misstatement penalties with non-uniform verification.