Which approaches avoid copying rows?
Explicit rules can construct cases directly from a schema and business requirements. Curated fictional fixtures can cover a small regression set. A learned model may require representative training data: confirm the origin and permissions for that input rather than assuming no production access.
Does synthetic mean anonymous?
No. Assess identifiability and linkage in the actual context. Pseudonymised data can remain personal data; learned outputs can reveal information about training records. The EDPB distinguishes pseudonymisation and anonymisation; its 2026 anonymisation guidance was issued for consultation.
What should the proof of concept show?
Trace inputs, models, reference files, logs and outputs. Use a business oracle and inspect whether rare and invalid cases are present. Document who can access each artifact, how it is deleted and which privacy and utility questions remain unverified.
Candidates for this task
- Faker: Generate individual values in test code; relationships and constraints remain your responsibility.
- DATAMIMIC CE: Define generated values and relationships in XML models; check that your model uses no sensitive source rows.
- GenRocket: Design generated scenarios for execution by its runtime; inspect the inputs and target workflow.
- Tonic Fabricate: Generate datasets from descriptions; do not upload real examples if the requirement forbids them.