The single most important operator rule: recall is not truth. A model's confidence is a property of its writing style, not of the facts. It will name a plausible book that doesn't exist, cite a version number from two releases ago, or invent a function that a library never had — fluently.
The discipline is boring and non-negotiable: anything that will be acted on gets verified against a live source. Not because models are useless, but because verification is cheap and acting on fiction is not.
The seed is a trap: it asks for details about something that changes constantly. Run it on two models and fact-check both against reality. The lesson isn't which model won — it's that neither gets your trust for free.