This lesson turns the evaluation habit into something small and permanent you can maintain yourself: a checklist of cases your product must always get right, no engineering required.

A must-pass case is simply a specific input paired with the answer your product must always produce for it. "When a user asks for their next appointment, it returns the correct one" is a must-pass case. You build the list straight from your Analyze step: take your most frequent, most important failures and turn each into a case that catches it. If "invents a detail that is not in the source" was a top failure, a must-pass case is an input where you know the correct answer and can check the product did not invent anything.

Keep every case binary, pass or fail, for the same reason as before: you want something you can check quickly and unambiguously, not a fuzzy score. And notice that cases come in two kinds. Some can be checked by a simple rule, does the output contain this, does it avoid that, which you can verify at a glance or with a trivial check. Others need a judgment call, is the tone right, is the summary faithful, which a person, or a validated AI judge, has to assess. Knowing which kind each case is tells you how to check it.

The discipline that makes this worth anything is re-running it. A checklist you write once and forget is useless; a checklist you run every time you change the product is a safety net that tells you immediately whether you improved things or broke them. That is the whole payoff, you can now change your product and know, in minutes, whether it still does the things it must. Start with a handful of cases covering your top failures, and grow the list as you find new ones.