META-GUARD: Public Reproducibility Package for Decision Robustness to Bibliographic Metadata Loss

Abstract

This public reproducibility package accompanies the manuscript “Decision Robustness to Bibliographic Metadata Loss in Automated Research Evaluation: A Paired External Evaluation Study.” The study used a 1,031-record development benchmark and a frozen 1,240-record external evaluation set comprising 900 controls and 340 positives. The archive contains locked study configurations, source and freeze manifests, dataset hashes, non-identifying aggregate result tables, perturbation and analysis code, numerical-audit and supplementary-table code, the frozen rule-engine implementation, software environment specifications, and programmatically generated figures. The complete internal test suite passed 18 of 18 tests, the EV1 data gate passed all prespecified checks, and the manuscript and supplement were numerically consistent with the locked aggregate results. Record-level bibliographic metadata, labels, provenance fields, per-record predictions, adjudication materials, API caches, and LLM response caches are excluded pending licensing and reputational-safeguard review. The public package supports verification of the reported denominators, aggregate results, principal conclusions, and all six figures. A complete record-level refit requires separately governed assets identified by the hashes preserved in this archive.