Check before you trust
A language model will hand you a sentence that's polished, confident, and wrong — and it won't blush. That isn't a rare edge case. It's the normal failure mode of the technology, which is why "check the work" from chapter zero is about to graduate from habit to discipline. This chapter turns it into evals.
So the question is never:
Did the answer sound good?
The question is:
What would prove this output did the job?
For a meeting-notes tool, proof might look like:
- every action item has an owner
- deadlines are copied from the notes, not invented
- uncertainty is flagged instead of hidden
- the follow-up email doesn't promise work nobody agreed to do
For a JSON extraction tool:
- the output parses
- required fields exist
- numbers are in range
- unknown fields are rejected
- examples that failed before stay fixed
That's the core of eval-driven AI development. You're not trying to make AI feel reliable — feelings are what fooled you in the first place. You're building a gate that catches unreliable output before it reaches a user.
A tool without checks is a demo. A tool with checks can become a system.