My understanding about your third point is the LLM generates lots of different answers, then they’re ranked according to some computation the creators came up with.

I’m still not sure what doesn’t qualify any of that as a prediction, and I’ll be more blunt: a guess.