Blinded expert assessment of action acceptability in synthetic clinical outputs from local large language models

Abstract All scenarios and model outputs in this study were synthetic. We evaluated expert-rated primary-action acceptability in a newly generated 44-scenario dataset, separate from a previously published technical benchmark. The generation design comprised 1584 candidate attempts; exclusion of one technically unsuccessful source configuration left 1320 eligible outputs. Four hundred outputs were randomly selected in Python and independently assessed by eight multidisciplinary reviewers, yielding 3200 ratings. Neither author rated outputs. An independent physician-researcher with more than 20 years of clinical experience defined scenario reference ranges before expert assessment. Output-level majority judgments were the primary reporting unit; agreement and uncertainty were examined separately. A majority judged the primary action acceptable for 338/400 outputs (84.5%; Wilson 95% confidence interval [CI], 80.6–87.7%), unacceptable or insufficiently urgent for 60/400 (15.0%; 95% CI 11.8–18.8%), with two tied outputs. Majority unsafe judgments occurred in 60/61 below-range outputs (98.4%) and 0/337 within-range outputs. Fleiss’ kappa was 0.492 and Gwet’s AC1 was 0.772; scenario-family bootstrap 95% CIs were 0.416–0.554 and 0.735–0.806, respectively. Uncertain ratings were reported and conservatively reclassified in sensitivity analyses. Below-range action was an observed marker of expert-rated concern in this sample. The findings support further study of a proposed human-in-the-loop audit, not clinical safety, patient benefit, or model rankings.