How do you catch a prompt that quietly got worse after a model update?

To catch prompts that quietly degrade after model updates, a Reddit user recommends keeping a small fixed set of real inputs with known-correct answers. Re-running these samples every couple of weeks and checking format, required fields, and labels reveals when a provider silently changes the underlying model.