A preprint scores 279 health-AI records, finding evaluation details were more complete than architecture, governance and reproducibility.