Structured Data Extraction Checklist (JSON/CSV)

Use this when an AI extracts structured fields (JSON, CSV, a table) from unstructured input like documents, emails, or freeform text. - Is the output actually valid — parses cleanly as JSON/CSV, no truncation, no trailing commentary mixed into the data? - Does every field match its expected type (a date field with an actual date, a number field with an actual number, not a string like "unknown" silently breaking downstream parsing)? - Are required fields always present, even when the source didn't clearly contain that information — check what happens on a missing-data case specifically, not just the clean examples. - When information is genuinely absent from the source, is it represented consistently (null, empty string, a specific sentinel) rather than the model guessing a plausible-looking fake value? - Do extracted values match the source exactly for anything that needs to (IDs, exact figures, names), rather than being "close enough" paraphrases? - If the schema has enums or fixed categories, does every extracted value actually belong to that set, or are there near-miss variants that will fail downstream matching? - Does it handle multiple instances correctly if the source contains more than one record (e.g. multiple line items), rather than only capturing the first? Test specifically with a messy, ambiguous, or incomplete real-world example, not just a clean one — that's where extraction quietly breaks.