Human and AI Bibliographic Screening of Brazilian Health Datasets — Corpus, Prompts, Decisions and Reproducible Analyses

This deposit accompanies the manuscript “Prospective Documentation of a Human and AI Bibliographic Screening Workflow for Brazilian Health Datasets”. It preserves the bibliographic corpus, archived screening prompts and notebooks, model decisions and rationale codes, human calibration and audit decisions, and a reproducible analysis package. The frozen corpus contains 962 unique bibliographic records. The 139 calibration records informed prompt development and are excluded from the primary performance analysis (823 records). Final human relevance decisions comprise 223 inclusions and 739 exclusions across the full corpus; the primary analysis includes 179 human inclusions. These are article-level screening decisions, not a catalogue of distinct datasets with verified final eligibility. Historical preparation, calibration and screening files are separated from the analysis package prepared on 22 October 2026. That package includes derived CSV inputs, code, saved results, provenance and verification files. Original archived spreadsheets and JSON outputs are supplied separately. Model outputs labelled “raw” are processed decision objects, not complete provider response envelopes. Rationale coding was introduced after calibration and before the full-corpus run, then applied as independent post-processing to already-produced rationales without changing screening decisions or the frozen prompt. The README documents the differing human-verification procedures and limits of the documentary reconstruction. The related deposit 10.5281/zenodo.20446831 contains the separate grey-literature portal corpus, frozen in May 2026; it is not the input corpus for this bibliographic evaluation.