The prompt
What the model is weighing right now
The model always does exactly this. A peaked distribution ("2 + 2 =") is what feeling certain looks like; a flat one ("the best programming language") is what opinion looks like; and the revenue prompt shows a third case, a smooth confident spread over numbers with no ground truth anywhere in it. That last one is "hallucination": not a malfunction, just sampling working as designed on a question the weights never contained. Confidence in the prose is a property of the sampling, not of the truth, which is why you verify load-bearing claims, and why your classmate's identical prompt gets a different answer.