I have a few prompts that behave reliably in production, but I’m concerned that a model update could change their output even if I haven’t touched the prompts. What’s a practical way to compare the new behavior with the version I’m using now?

I’m especially interested in model snapshot comparison: do you keep a fixed test set and score outputs automatically, or review a sample by hand? If you’ve caught a subtle regression this way, what did you measure besides exact wording?