[mistermann]

This seems right to me as well: it abstracts the text to a temporal/ spatial object model, and the "What do you think about the following text?" prompt invokes analysis of that model according to prior training examples (as opposed to someone else above prompting it to describe the schedule, in which case it isn't concerned about the logic of the situation).

I would guess that the human mind does this abstraction behind the scenes invisibly, screwing up our intuition when analyzing how LLM's work. I wonder if using examples that are counterintuitive to human intuition might offer insight, because humans reveal their perceived logical thinking is not actually that (rather, is heuristics) in their post-hoc rationalization of the "logic" they believe their mind executed to produce the answer.

(I don't think I articulated what I'm thinking here very well...or, perhaps I have fallen victim to my very own theory!)

A bit more effort...the text is converted into not only tokens, but also abstract tokens, and it is because of the translation into abstract tokens that it is able to match it to training data (which would also have to be translated into abstract tokens). How it resolves the inconsistency after that translation though is beyond me, but it wouldn't surprise me if it is (in this case) a rather trivial problem to someone with depth in logic or some other related discipline.