CRLN-HB-001 — The Human Baseline for Clinical Research AI (Version 0.3, corrected)

This version corrects v0.2 and reverses its headline conclusion. v0.2 should not be cited for any statistic. CRLN-HB-001 measures how competent humans perform on scenario-based clinical research judgement items, so that an AI system's score on the same items can be interpreted. Items are anchored to ICH E6(R3) and CRLN-TRG-001, and are scored on a four-point behavioural rubric. The correction. The v0.2 baseline excluded two founder accounts and nothing else. CRLN flags 65 internal test accounts, and in the window v0.2 reports, 38 of 74 contributors and approximately 70% of scored decisions belonged to them. The published "human baseline" was mostly instrumentation. Corrected sample (identical cutoff, exclusion added): items at n≥5 fall from 129 to 28; human responses from 1,215 to 205; distinct humans from 57 to 35; weighted mean rubric rises from 3.641 to 3.810. The conclusion reverses. CRLN-AIE-001 (DOI 10.5281/zenodo.22708752) reported nine frontier models at 3.632 against humans at 3.641 and concluded the instrument could not separate them. Recomputing both arms over the same 28 surviving items by the same method gives humans 3.810 (n=205) against models 3.606 (n=938): a difference of 0.204, 95% CI 0.111 to 0.297, Welch t = 4.31, Cohen's d = 0.37. Competent humans outperform frontier models on these items by a margin well outside sampling noise. Stated limitation. The published model mean of 3.632 could not be exactly reproduced; reconstruction from stored responses gives 3.670 across all items. The comparison is internally consistent but is not a reproduction of the original scoring pipeline, so the direction and approximate magnitude are established while the exact model figure remains provisional. The corrected human figure does not carry that caveat. A generalisable failure mode. The document devotes a section to why no test or review would have caught this: the test-account flag lived in a different database from the code that needed it, so the property was not expressible where the work happened. The exclusion list looked deliberate, the contaminated data looked realistic, and the error flattered nobody — test accounts scored lower than real learners, so the contamination manufactured a more surprising and more publishable finding than the truth. Errors that make a result more interesting are the hardest to notice. Licensed CC BY 4.0.