ClinEvalclinical LLM evaluation

Equity

Unsupported-claim-rate and planted-fact recall across demographic and complexity slices. Toggle the metric; error bars show 95% bootstrap CIs.

Largest gap · Language profile
6.9 ptsnon-overlapping 95% CIs

non native english 9.7% vs interpreter assisted 2.8% unsupported claims. Their 95% CIs don't overlap, so the gap is unlikely to be noise.

Language profile

Age band

Sex

Complexity

Error bars show the 95% bootstrap confidence interval (1,000 resamples over whole dialogues). Overlapping intervals mean a gap is not robust.