Guardrails That Run on Every Request

A customer reports a dripping kitchen tap, and the job ticket assistant replies: "A plumber will be at your home by 3 pm today." The reply breaks a promise, because the assistant must never promise an arrival time.

Suppose the reply appears in the offline test set. A code check marks it as a fail, and a developer reads it the next morning and fixes the prompt. No customer ever sees it.

Now suppose the same reply is written for a real customer. An offline eval cannot help that customer, because it runs on test cases, not on live requests