Reproducibility Materials for Retrospective Evaluation of Gemini 2.5 Pro for Title–Abstract Screening Across Four Systematic Reviews

This dataset accompanies the manuscript Retrospective Evaluation of Gemini 2.5 Pro for Title–Abstract Screening Across Four Systematic Reviews and provides the materials required to inspect and reproduce its reported screening-performance estimates. Contents. The deposited package contains exactly two analyses for each of four systematic-review corpora (Original and Revised): eight prompt files, eight deidentified record-level decision files, eight harmonized run summaries, aggregate complete-case and missing-abstract metric tables, a field dictionary, reconstruction and metric-calculation programs, and checksum manifests. The decision files comprise 36,226 rows. The clean 1,807-record canonical-prompt analysis is the sole R2 Original analysis. Interpretive scope. Revised prompts were developed after examination of Original disagreements and were applied to the same corpora. In R1–R3, the Revised prompts also changed parts of the operational eligibility target. The paired estimates therefore describe two completed screening versions and should not be interpreted as independent same-target validation estimates. Privacy and redistribution. The public decision files use opaque record identifiers and contain no article titles, abstracts, author names, journal names, bibliographic identifiers, or raw model-response identifiers. The underlying title–abstract libraries are not redistributed because they contain third-party bibliographic records and materials subject to pre-publication restrictions. Verification. After extracting the archive, run python verify_reproducibility.py --root . from the Reproducibility_materials directory. The metric-calculation program independently reconstructs complete-case and missing-abstract estimates, Wilson 95% confidence intervals, and 100,000-replicate bootstrap intervals for work saved over sampling. The package uses only the Python standard library. Related registration. The study registration is available from OSF at https://doi.org/10.17605/OSF.IO/HDSV4.