Shelf five · Putting it to work
Evals & Hallucination
Kitchen table
You keep a secret stack of questions you already know the answers to, and you make the machine sit the same test every week. Not because you enjoy marking, because machines can get worse without telling you. An update meant to make it politer might quietly make it worse at maths. The test stack is how you notice. And when the machine invents an answer that sounds right but is not. That has a name, hallucination, and it is why you never mark its homework by vibes.
Whiteboard
Task-level suites, graders (exact-match where you can, LLM-as-judge where you must, with the judge itself evaluated), golden sets guarded from contamination, and regression runs wired into CI so a prompt change cannot silently ship a capability loss. Writing down what 'correct' means is the uncomfortable part and the entire point: most teams discover they never agreed. Evals are to AI what tests are to code, with one difference. The system underneath drifts even when you change nothing.
Boardroom
One question separates serious AI vendors from the rest: how do you measure that it is right, and can I see the test set? Vibes are not a control, and a demo is a performance, not a measurement. Internally, insist every AI feature ships with its eval suite and that the scores are reported over time like any other operational metric, because the failure mode is not day-one wrongness, it is quiet degradation while everyone assumes someone else is watching. The eval budget is the insurance premium.
Related concepts