Dataset for: Source-Identifier Leakage and Gold-Standard Ambiguity Inflate Measured Recall in Published-Reference Recovery Benchmarks for AI-Assisted Evidence Synthesis

Data and code release supporting the MARS-B manuscript. This is the corrected version, aligned with the article's measurement-validity framing (it replaces the original 3 June 2026 deposit; same concept DOI). Contents: the per-chain × per-model recovery matrix, the locked rerank prompt and output schema, the 16 production run-output logs and one excluded-sensitivity run, derived analysis outputs, the TRIPOD-LLM checklist, analysis/prefilter/audit scripts, and validation manifests (MANIFEST + SHA256SUMS). To respect source-document copyright, the deposit excludes full-text sources, candidate-CMO text, embeddings, and raw transcripts (see EXCLUDED_MATERIALS.md). Code under MIT; derived analytic data under CC BY 4.0.