You’ve deployed your LLM. Prompt engineering is solid. The RAG pipeline works in staging. Then production hits — and the model starts hallucinating like ...