Generate or augment training and test data while measuring realism, coverage, privacy leakage, bias, and whether synthetic examples improve real-world performance.