What changes the decision?
Statistical similarity, task utility and privacy are separate properties. A dataset can resemble a source and still fail downstream evaluation. Seed datasets steer generation; they do not establish a random-seed replay contract.
What should your proof of concept show?
Keep a real holdout outside training. Compare task performance, rare groups and constraints, then check disclosure or memorization risk. Ask for a versioned evaluation report rather than a single synthetic-data quality score.
These are editorial decision criteria, not measured product rankings. Follow the profiles below for product-specific source evidence.