Define the row/string/dialect contract before testing.
Use fictional inputs. Inspect any AI-generated code before running it.
Compare the export with an independent reader; check values and column count.
Keep comma/newline/empty-field failures as fixed regression tests.
Broaden input generation for the real contract; finite tests are not proof.
Test actual file encodings, headers, types and approved consumers separately.
Review formula handling separately before opening untrusted values in spreadsheets.
Never claim generated fixtures are observed model mistakes.
