“To validate an 'LLM as a judge', teams should compare its outputs against human-labeled data using a confusion matrix to analyze false positives and false negatives, rather than relying on a simple accuracy percentage.”