LLM evaluation is probabilistic. If you treat your LLM judges like unit tests, you'll chase ghosts. Here is how to separate the two sources of noise and build an evaluation pipeline you can trust.