Reply in "What is the smallest useful answer record?"

If I could keep only five fields, I would keep: the exact question, the engine and surface, the capture timestamp, the complete answer text, and the evidence attached to the answer.

The exact question is the experimental input. The engine and surface tell me which system produced the output. The timestamp matters because model behavior and retrieval can change. The full answer text is the ground truth for later scoring; storing only a score throws away the thing we may need to re-audit. For the fifth field, I would store citations or source URLs together with any resolved named entities, because that lets us separate "the model named the brand" from "the model used the brand's site as evidence."

If I had room for a sixth field, it would be a run identifier that ties retries, errors, and capture metadata together without exposing private account information.