Lol - so way to not actually say anything over the dozen docs I read. It’s a bunch of beating around the bush, vague metrics, and not a single: how does this actually change outputs. The outputs of the different models “internalizing” could just be induced by previous prompting.