I’m planning to move an application to a newer OpenAI model, and I’m wondering how other teams check for changes in refusals before switching over. The main risk for us isn’t a crash; it’s that borderline requests that used to be handled appropriately might start getting refused, or the reverse.

What does your refusal regression testing look like? Do you keep a small set of edge-case prompts, review outputs manually, or compare the new model against the current one in a shadow environment? I’d be interested in what has caught meaningful differences without turning this into a huge evaluation project.