Learn
How to validate synthetic data
Vendor-neutral explainers on the methods behind a validation report. Every number quoted comes from a real run you can inspect at /sample.
Start here
Synthetic data validation: the complete guide
How to validate synthetic data across fidelity, privacy and utility, why one can't offset another, and how to read a report.
Metrics
One page per measurement: what it means, how it is computed, and what a bad score usually indicates.
Column shape: KS statistic and total variation distance
How per-column fidelity is measured with KS and total variation distance, what a bad score means, and why passing it isn't enough.
Correlation preservation
How pairwise fidelity is measured, why collapsed and invented correlations are both findings, and how it connects to utility.
Distance to closest record (DCR)
What DCR measures, how it detects memorisation in synthetic data, why it needs a baseline, and how to read it with NNDR.
Exact-copy detection
The simplest privacy check: how many synthetic rows are real rows. Why it always runs, what near-copies mean, and why zero copies isn't a pass.
Membership inference
What a membership inference attack measures, why it needs a holdout, what a bad result means, and why 'not tested' isn't 'passed'.
Nearest-neighbour distance ratio (NNDR)
What NNDR measures, how it flags synthetic rows that hug one real record, and how to read it with DCR.
Record isolation, linkage and inference risk
The EDPB's three anonymisation criteria (formerly singling-out, linkability, inference) as attacks on synthetic data — naive vs residual risk.
Train on synthetic, test on real (TSTR)
What TSTR measures, how it's computed, what a bad score usually means, and how to read it alongside TRTR.
The holdout question
Why attack-based privacy metrics need real records the generator never saw.
Benchmarks
The explainers describe the method; these are the method applied. Real reports on public datasets, published unedited — the evidence behind everything above.
Compliance
For DPOs and privacy engineers: what a validation report is evidence of, and what it is not.
Is synthetic data personal data under GDPR?
What GDPR and the EDPB's 2026 draft anonymisation guidelines say about synthetic data, the three criteria, and what evidence a controller needs.
Synthetic data in a DPIA: what to document and why
How synthetic data fits into a Data Protection Impact Assessment, what evidence belongs in it, and how to phrase claims honestly.