I’m testing a customer-support workflow with the OpenAI API, and I’ve seen an occasional refusal for requests that look ordinary and are nearly identical to ones that worked moments earlier. I’m trying to work out whether I’m looking at a safety-policy decision, something in the surrounding conversation, or just a difference in the generated response.

For anyone who has investigated this in production, what request metadata or response details were useful to compare? I’m especially interested in how you distinguish a genuine policy refusal from a response that simply failed to follow the prompt.