Most LLM evaluation failures trace back to a handful of structural errors: contaminated test sets, unanchored graders, and prompt drift. This walkthrough diagnoses each one against a real evaluation pipeline.