Synthetic data is never trusted by assumption. Evaluation stress-tests every generated batch with 20+ metrics, so you know — not hope — that the augmentation matches reality.
The five report families
Quality Report — Distribution Match Score for continuous columns, Category Match Score for categorical, plus column-pair correlation similarity.
ML Efficacy — an Indistinguishability Score measuring how easily a classifier can tell real rows from synthetic ones. The harder that is, the more faithful the data synthesis.
Dimensionality reduction — Linear Map and Structure Map projections of real vs. synthetic rows, so divergence is visible rather than just scored.
Privacy Shield — record-level distance to the closest real row, an overfitting guard, and an exact-duplicate check. Relevant the moment synthetic data is shared outside your lab.
Anomaly detection — unsupervised, threshold-free detection of the worst-fitting rows, flagged for manual review.
Reading the headline scores
Overall Quality Score — the summary: how close the synthetic distribution is to the real one.
Correlation Similarity — whether relationships between columns survived generation.
Distribution Match Score — per-column distributional accuracy.
What to do with the numbers
High quality + high correlation similarity means the synthetic data is a safe stand-in for the real set in modeling. A low Indistinguishability Score with a weak quality report means the generator struggled — switch modes (try Statistical on small data), reduce the number of rows generated, or fix mislabeled column types before regenerating.