Your eval suite is green. Production is not.

You ship an LLM feature. Three weeks later a Slack thread mentions that a customer got a strange answer. You dig in. The prompt has been edited twice since release, the provider has quietly re-quantis