Utility Benchmarks

What To Measure

Utility should be measured against the downstream decision, not only a generic score.

Examples:

• Model accuracy by site, class, and subgroup
• Synthetic data utility for specific analytic tasks
• Retrieval quality under access constraints
• Aggregate measurement error under DP noise
• False positives and false negatives in PSI-backed workflows

Reporting

Report baselines, variance, subgroup behavior, and cases where the PET changes the decision.