Asking for the Whole Plan: Prompt Design Moves Language Models' Scores on a Clinical-Safety Benchmark More Than Model Choice. Data and code

Data and code for the manuscript "Asking for the Whole Plan: Prompt Design Moves Language Models' Scores on a Clinical-Safety Benchmark More Than Model Choice. An Exploratory Study on the NOHARM v2 Public Cases" by Ahmad Omar Aljamal, MD (preprint: https://www.preprints.org/manuscript/202610.1607). This is a snapshot of https://github.com/OJamals/noharm-prompting at commit 0b2b800, which carries version 2 of the manuscript. Contents: every model plan and AI evaluation for the 30 public NOHARM v2 cases (330 items), the off-rubric action ratings, the physician review materials and item-by-item verdicts, every instruction given to a model, the scripts that ran the arms, and the offline generators that reproduce every number in the paper without model calls. The manuscript (version 2) and its Supplementary Appendix are included as LaTeX and PDF. Files: the repository is split into five zip archives because of upload size limits. Unzip all five into the same folder to rebuild the repository tree. Part 1 holds the code, documents and small datasets; parts 2 to 5 hold the raw model results under sources/mast/results. Scores come from a locally run AI evaluator, not from the benchmark organizers, and nothing here is a clinical validation. Licenses: code MIT; data produced in this study CC BY 4.0. The NOHARM kit (cases, rubrics, metric code) is Copyright (c) 2026 ARISE Network, code MIT and data CC BY 4.0; cite the NOHARM authors when using the cases or rubrics.