From an email I just sent:
My view was this: if you play out the main computer lines for 8-10 moves, Black’s position does not really improve, nor are White’s holding moves hard to find (he just has to shuffle back and forth).
Black does not have the structural advantage to enable a later transposition into a favorable endgame.
So it actually is a draw! (sort of)
Does the agreement[to draw] have to be non-Bayesian? There is a “vague range,” and by Magnus Carlsen offering the draw maybe he, as Kasparov suggested, lowered his own chances for the rapid tiebreak (shows some loss of nerve), until they were in the same “vague range” as the game 12 final setting.
So Magnus is saying “I don’t have a way of pressing that is better than my chances in the rapid tiebreak,” and Caruana is agreeing and knows that Magnus knows this.
Maybe not “strictly” Bayesian, but it doesn’t seem crazy to me either.
I thank S. for a relevant conversation on these points.