FIELD GUIDE 01

Define the synthetic-data task

Choose synthetic data for the task it must perform.

Test data and model-training data are not interchangeable. Define the schema, fidelity requirements and privacy controls, then validate the generated output against its downstream use.

Define output and task fit.

Write down the current process, the expected output and the person responsible for approving it. For data engineering, make output and task fit an explicit part of the brief rather than an assumption hidden in a demo.

Give each provider the same task.

• Output and task fit: state the requirement, ask for evidence and record any limitation before the pilot.
• Fidelity tests: state the requirement, ask for evidence and record any limitation before the pilot.
• Privacy controls: state the requirement, ask for evidence and record any limitation before the pilot.
• Integration and reproducibility: state the requirement, ask for evidence and record any limitation before the pilot.

Questions for the shortlist.

Tonic: Choose the specific Tonic product: generation, structured masking and text de-identification have different pricing. Read the official product description ↗

MOSTLY AI: Useful when the evaluation should include local compute and programmatic generation. Read the official product description ↗

DataMasque: A different starting point when you already have source data to protect rather than invent from scratch. Read the official product description ↗

Check the exception path.

Include missing inputs, ambiguous requests and a failed integration in the evaluation. Record the result and the recovery steps. A system that works for the routine case still needs an accountable route for the rest of the work.

Decide what would justify proceeding.

Agree an acceptance threshold for fidelity tests with the people who operate the process. Keep the input set, result notes and unresolved questions so the decision can be revisited when the product or requirements change.