How training a tiny transformer model helped LLM hallucinations feel less like magic and more like a fundamental reliability problem in AI.