Pick something you know well. The AI will write eight confident claims about it. Some are true. Some it made up on the spot. You get ten seconds a claim to say which.
Topics only. Not people.
Nothing you type is stored. The topic goes to the AI, the claims come back, and that's the whole transaction.
Sample round: this copy of the game is running on a fixed set of claims instead of the live AI.
Then a second, separate AI reads them cold and rates how sure it is each one is true. You'll see those numbers after you've made your calls.
The confidence numbers come from a second AI that saw the claims without labels. It had no way to know which ones the first AI invented. That's the point: confident and correct are different things, and the only check that works is knowing the subject.