LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It
Sebastian Fox, Luke Markham, Ryan Lail, Michael Karotsieris
cs.CL · cs.AI
Quick verdict
Claims LLM judges miss omissions in AI clinical notes near chance; benchmark and ablations strongly support it; remedies only partly recover detection.
Summary
The paper asks whether LLM judges—the standard automated quality check for AI-generated clinical notes—can detect omissions. Public corpora cannot supply an answer key because clinician reference notes are discrepant with their own transcripts (all 53 audited PriMock57 and 45 ACI-Bench notes carry material discrepancies), so the authors build a 500-pair benchmark from transcript-derived, audited fact sheets: 298 omission pairs (complete, fragment, restatement-trace) and 202 added-or-altered controls. Across eight ablated judge designs, paired discrimination reads 0.79-0.94 on commissions but 0.50-0.63 on omissions; on single notes no design separates omission notes from clean ones at usable false-alarm rates. Five remedy families (criteria, answer format, voting, wording, GEPA prompt optimization) shift the operating point without creating usable detection. What works is restructuring: a per-fact pipeline (24.6% detection at 2.7% false alarms) and a budgeted GEPA-evolved prompt (36.9% at 6.2%), the latter detecting more (p=0.002) at a tenth the cost ($0.046 vs $0.45 per note). A physician author validated 70 items; an independent clinician's blind grading of the severity rubric agreed within one grade. Limitations: no benchmark threshold transfers to 261 real vendor notes; restatement-trace omissions defeat all methods; the physician validator is an author; the pipeline shares design lineage with benchmark construction. Benchmark, prompts, and judgements are released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Across eight ablated judge designs and every reference judge but one, LLM judges separate commissions from clean twins (0.792–0.944 paired) but sit at 0.500–0.634 on omissions, with no usable single-note detection at deployable false-alarm rates.
- Five remedy families — criterion scope, answer format, voting, wording, and GEPA prompt optimisation at the tested budgets — move the operating point without creating usable omission detection.
- Converting the absence question into presence checks recovers partial detection: a per-fact pipeline reaches 24.6% detection at 2.7% false alarms and a budgeted GEPA-evolved prompt reaches 36.9% at 6.2%, detecting significantly more (p=0.002), with closed per-fact verdicts supplying most of the gain.
- No benchmark operating point transfers to 261 real vendor notes without recalibration, and restatement-trace omissions defeat every deployable method tested.
Strengths
Absence is made certain, not assumed
Construction verifies removals: only 71/126 complete-removal attempts survive a cross-family panel (56.3%), 92 pairs are dropped and none patched (§3.2, Fig 2); blind extraction recovers 646/650 (99.4%) of facts human authors wrote for the same consultations (§3.2, Appendix A.3).
The asymmetry is isolated by design, not by one prompt
Eight ablated designs (Table 4) read 0.792–0.944 paired on commissions vs 0.500–0.634 on omissions; the pattern replicates in reference judges (§5.3), a Gemini-family judge (0.951/0.961 vs 0.551/0.683), and the commission arm is externally anchored on physician-labelled MEDEC (AUC 0.811 vs 0.575, §5.4).
Null results are quantified, not asserted
Mechanism is located with controls
Transfer failure is reported against the authors' own interest
Weaknesses
The pivotal human validation is author-run
Construction–judging circularity is mitigated, not eliminated
Asymmetric rule-setting favours the proposed methods
Clean twins bound the false-alarm metric optimistically
Complete omissions are a construction-survivor minority
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether the LLM judges used to quality-check ambient AI scribe output can detect omissions — facts the consultation transcript establishes that the note fails to record. Because audited public reference notes prove unusable as an answer key (all 53 retained PriMock57 and all 45 ACI-Bench notes carry material discrepancies with their own transcripts, at means of 10.7 and 7.1 per note; §3.1), the authors construct a 500-pair benchmark (495 in the evaluation set) from transcript-derived fact sheets: 298 omission pairs — complete, fragment, and restatement-trace — graded by severity and surviving trace, plus 202 added-or-altered controls. Central claims: (1) judges verify presence, not absence — 0.792–0.944 paired discrimination on commissions versus 0.500–0.634 on omissions across eight ablated designs (Table 4); (2) five remedy families shift the operating point without usable detection; (3) enumerate-then-check restructuring recovers partial detection via two independently derived implementations.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is omission blindness in LLM judges?
The paper's finding that LLM judges reliably catch content wrongly added to or altered in a clinical note (paired discrimination 0.79–0.94) but perform near chance on content the source transcript established yet the note fails to record (0.50–0.63), because an omission leaves no text for an evidence-seeking judge to point at.
What fix works for detecting omissions in AI clinical notes?
Replace the single open question 'is anything missing?' with closed presence checks: list the facts the transcript establishes, then verify each against the note. A per-fact pipeline detects 24.6% of omission notes at 2.7% false alarms; a GEPA-evolved single prompt run with a reasoning budget detects 36.9% at 6.2% false alarms and costs about a tenth as much per note.
Is the OmissionBench dataset public?
Yes. The 500 graded note pairs, fact sheets, severity grades, prompts (including superseded versions), model pins and every benchmark judgement are released at huggingface.co/datasets/ComposoAI/OmissionBench under CC BY 4.0, with the harness on GitHub (composo-ai/omission-bench) under MIT and archived on Zenodo. Vendor note text is withheld for licence reasons.
Can these judges be trusted on real clinical notes today?
Only partially. On 261 real vendor notes every benchmark threshold moved — false alarms rose to 22.4% (pipeline) and 44.2% (evolved prompt) — so thresholds must be re-established per deployment, and even the best method still misses roughly two-thirds of notes containing an omission.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with several [...sampling gap...] breaks; mid-paper regions (parts of §6, §7.4–7.5, and appendices B–K) were skipped, so judgments touching those regions rest on cross-references that were visible in the retained text.
- The severity rubric verbatim text, the GEPA campaign prompts, and the Appendix K transfer protocol were not fully visible in the excerpt; their quality is assessed only via in-text summaries and tables.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.