“Goodfire research showed it is possible to distinguish model weights used for memorizing facts from those used for general-purpose reasoning.”