I’m looking at using the OpenAI Batch API for a large set of independent requests, but I’m unsure how best to pair each result with its original input. If the output file isn’t in the same order as my submitted requests, are the custom_id values the dependable way to do that?

How do you handle missing, failed, or duplicated results when processing the batch? I’d rather not assume the ordering is stable and discover later that a response was attached to the wrong record.