How should you evaluate learned synthetic data?

What changes the decision?

Statistical similarity, task utility and disclosure risk need separate checks. A dataset can resemble its source and still fail a downstream task. Small rare groups may be the cases that matter most. Conditioning rows and a random seed are different inputs; neither alone guarantees replay.

What should your proof of concept show?

Keep a real holdout outside training. Set task metrics, subgroup checks and tolerances before fitting. Check constraints and disclosure risk separately. Record training inputs, model, software versions and sampling settings, then repeat the experiment.