What should we check after transforming review data for our marketplace?

Run these checks on the customer-facing export after every transformation that can change record meaning.

- Preserve identifiers as strings, including leading zeros. - Keep unknown values distinct from negative answers. - Retain all members of nested evidence lists. - Resolve every required cross-file record reference. - Preserve policy text without silent truncation. - Keep decision labels aligned with their review records. - Compare reconstructed tasks with the source task contract.

A marketplace may flatten nested JSONL into tabular columns or convert records into Parquet for customer downloads. Those operations can introduce defects even when the generated source is coherent. An empty evidence list, an omitted field and an explicit statement that no evidence exists can mean different things to a review agent.

Use a small set of deliberately varied records to exercise those distinctions. Include optional fields, multiple findings and records with no final decision yet. Validate both machine parsing and the task presented to a customer.

A useful acceptance question is whether a downstream user can reconstruct the same review problem from the exported files. File validity alone is insufficient when the conversion has silently removed an evidence relationship or replaced an unresolved state with a default.