ROP-LLM Clinical Decision-Safety Benchmark and Reproducibility Package

This dataset and reproducibility package accompanies a multimodel clinical decision-safety evaluation of large language models in retinopathy of prematurity (ROP). It includes 500 structured simulated ROP scenarios with locked expert reference standards, verbatim Standard and Safety-focused prompt templates, counterfactual and repeatability scenario sets, the locked 5,250-output analysis workbook, machine-readable statistical outputs, reproducible analysis and validation code, software records, audit files, and supplementary reporting tables. The evaluated task begins after retinal examination findings have been documented and assesses translation of those findings into urgent-treatment, follow-up, or screening-termination decisions. No real patient data, medical records, clinical images, or human-participant data are included.