OmissionBench: a benchmark and harness for omission blindness in LLM judges of AI clinical notes
Year: 2026
Type: software
Abstract: OmissionBench is the evaluation substrate behind two companion papers on whether automated judges detect that a clinical note has left something out. This record archives the tagged harness release (v1.0.2). The data is released separately at huggingface.co/datasets/ComposoAI/OmissionBench under CC BY 4.0; the code lives at github.com/composo-ai/omission-bench under MIT. The benchmark is 500 single-error clinical note pairs built from audited, transcript-derived fact sheets, graded by how much a removed fact matters clinically and by how much of it survives elsewhere in the note, together with the verified findings of a census of three deployed commercial ambient scribes. The three scribe products are anonymised throughout and their note text is not released. This is an archival research release rather than a maintained project.
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204717840