State the origin of every subset
Require clear labels for operational records, staged examples, and synthetic material. Explain how synthetic documents were produced and whether authentic templates or values informed them. Mixing origins without labels makes it difficult to determine whether strong performance reflects operational competence or familiarity with a generator.
Use controlled variation intentionally
Synthetic examples can help exercise specific cases such as long tables, missing fields, or rare formatting combinations. Specify the intended stress condition and expected output. Avoid assuming that a generated exception has the same frequency or contextual complexity as the corresponding event in an operating business.
Keep a realistic evaluation boundary
When deployment involves authentic documents, reserve an appropriately reviewed authentic evaluation slice where feasible. Group related templates and generators so near-identical versions do not cross splits. Report performance by origin and difficulty instead of averaging all records into one number that hides weaknesses.
Make the brief transparent
Tell Fiund whether authentic records are essential, whether synthetic supplements are acceptable, and which sources need independent review. Request counts and examples by origin. Define the supporting provenance documentation and allowed transformations before sample selection, including how sensitive values should be handled in each subset. A controlled comparison can use the same field schema across authentic and generated subsets while retaining their separate identities. Inspect which errors occur only on authentic examples, such as ambiguous stamps or process-specific annotations. Those differences can reveal where additional synthetic variation would still fail to address the actual operational challenge.
Review checklist
- Label origin at record level
- Identify the purpose of synthetic cases
- Report evaluation by source origin
- Specify authenticity and review requirements