Where Reproducibility Policy Actually Lives: A Stratified Audit of 150 Clinical Journals — Data and Code

Le résumé fourni par la source

Responsible use and dual-use — read before reuse This repository releases code and data that measure the machine reachability and legibility of clinical-journal editorial policy at scale. It is released for reproducibility, legitimate meta-research, and accountability. The access friction this work documents — bot-blocked policy pages, navigational depth, non-machine-readable formats — is not a safety control. It is an accidental by-product of publisher infrastructure that obstructs legitimate audit indiscriminately. Legibility is dual-use: the same reachability that lets an outside audit read editorial policy at scale also lowers the barrier to autonomous machine operation over the scholarly record. Because the current illegibility is accidental rather than a deliberate safeguard, it must not be relied on as one; opening it obligates building deliberate controls in its place — provenance, audit logging, and access tied to accountable, identifiable use — rather than leaving friction to do that work by accident. We advocate machine-legible policy for meta-research and accountability, coupled to those controls, not as an unconditional opening of the record. See §4.1 of the accompanying paper. Open data and code for a cross-sectional audit of the documented editorial infrastructure to evaluate computational research across a field-normalised (Scopus SNIP) stratified random sample of 150 MEDLINE-indexed clinical journals in six specialties (general internal medicine, cardiology, infectious diseases, oncology, neurology, pediatrics). Five indicators — statistics/methods editor, AI/machine-learning editor, data-availability statement, code deposit, and executable artifact — were hand-coded from public editorial-board and author-instruction pages, separating disclosure (a statement that data exist), deposit (a requirement to share), and verification (a check that a computation runs). Codes were verified against a timestamped source-page archive; the data-statement and board codes were cross-checked by blind automated agents; and a boundary-enriched 50-cell subset was re-coded by an independent second coder (Cohen's κ=0.69, three of five indicators). Contents: Frozen sampling frame and draw order (600 drawn, 150 retained) Per-journal coded indicators with source URLs Publisher-family policy units (149 readable author-policy records → 50 analytic units within the 150-journal sample) Three-profile HTTP reachability probe results and the automated-comparator (machine-vs-human) codes Codebook, coding rules, and analysis.py, which regenerates the reported prevalences, intervals, statistics, Tables 1-4 and Table S1 (make_figures.py regenerates Figs 1-3; Figs 2-3 are data-computed and Fig 1 is a schematic; the PubMed trend is not regenerated) Journal-level policy is reported openly as a matter of public record; editorial roles are coded for presence only and no individual editor is named.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.