Local LLMs fail differently than their cloud counterparts. Here is a concrete debugging methodology for Llama and Mistral deployments, built around one in-depth case study.