LLM Judges Verify Presence, Not Absence: Blind Spot in AI Clinical Notes

Ambient AI scribes draft clinical notes from patient encounters, but published audits show the dominant error is omission — information the encounter established that never made it into the note. The standard safeguard is an LLM judge: a second model reads the note against the transcript to flag problems. This study asks whether judges actually catch omissions.

Using a controlled benchmark of 500 single-error note pairs drawn from audited fact sheets (298 with a genuinely missing fact, 202 added-or-altered controls), the researchers tested eight judge designs. Discrimination on added or altered content was strong (0.79-0.94), but on omissions it dropped to 0.50-0.63 — barely better than chance. Prompt rewording, voting, and GEPA prompt optimization failed to fix this gap.

The fix comes from restructuring the task: first enumerate the facts established by the transcript, then check the note against each one. Two independent methods achieve this recovery — a per-fact pipeline that names the missing fact and its severity at a 2.7% false-alarm rate, and a single-call prompt evolved via GEPA that detects more (36.9% vs 24.6%) at a 6.2% false-alarm rate and a tenth of the cost. A physician reviewer sided with the pipeline in all 10 disagreement cases tested.

On real-world vendor notes, benchmark thresholds don't transfer directly, but the recalibrated single-call method still outperforms the best of the eight original judges at half the false-alarm rate. Omissions restated elsewhere in the note remain a blind spot for both approaches. The benchmark, prompts, and judgments have been released publicly.