Reproducibility bundle: Iterative Prompt Calibration of a Vision-Language Model System for Automated Radiology Report Generation

Companion code and label-level data for the arXiv "Iterative Prompt Calibration of a Vision-Language Model System for Automated Radiology Report Generation" (M. Kozłowski, 2026). The bundle reproduces every Study 2 result (Tables 3-4, Cochran Q, McNemar, per-stratum and case-level statistics, and all Study 2 figures) from the released label-level annotations, without requiring API access. It contains: the Study 2 evaluation pipeline (multi-model runner, Claude-Haiku label extractor, scoring and analysis scripts), the full System A/B/C prompt texts, the finding-level classification file (annotation_auto.json), and the exact analysis outputs (stats.json, summary, LaTeX tables). MIMIC-CXR images and raw model-generated reports are NOT included (PhysioNet credentialed-access data use agreement); the Study 1 clinical DICOM cases are not shared for patient data-protection reasons. See README.md for reproduction steps and the mapping to paper tables/figures.