LLM judges score 0.50-0.63 on detecting omissions against 0.79-0.94 on detecting additions.

Evaluating AI-generated clinical notes, judges verify presence but not absence, and no single-note judge design flagged omissions more reliably than it flagged perfect notes (arXiv 2608.31016). Restructuring to first enumerate transcript facts then verify each against the note raised omission detection to 24.6-36.9% at 2.7-6.2% false alarms, with a physician reviewer siding with the pipeline on 10 of 10 disagreements (p=0.002). If you're judging summarization output, your judge is blind to the failure mode that matters most.