Dataset for: From peer-reviewed realist synthesis to published-reference benchmark — source-paper recovery and large-language-model rerank portability in the Aunger/Millar context-mechanism-outcome corpus

Data and code companion to the manuscript "From peer-reviewed realist synthesis to published-reference benchmark: source-paper recovery and large-language-model rerank portability in the Aunger/Millar context-mechanism-outcome corpus." Contains the per-chain × per-model recovery JSONL (Additional file 2; 16 production runs × 328 chains; 291 Robust-eligible units), the pre-submission sensitivity-analysis output (Additional file 3), the TRIPOD-LLM checklist (Additional file 4), the supplementary appendix (Additional file 1), the locked rerank prompt and output schema, the prefilter and analysis scripts, the per-model run logs, and the deposit-validation manifests. Full-text source documents, PDFs, abstracts, embeddings, and candidate CMO text are excluded to avoid redistributing copyrighted source material; only metadata, derived analytic outputs, and pointers to canonical sources are included. Rendered figures and their plotting scripts are omitted from this minimal data/code release and are available from the corresponding author on reasonable request.