Year: 2026
Type: preprint
Abstract: Research software citation metadata was written for human readers and patient indexers. Its audience has changed. AI agents now read research software's self-description directly from the files a project publishes about itself: CITATION.cff, codemeta.json, zenodo.json, README citation sections, archival DOI records, and package-registry entries [1-7]. These consumers do not ask clarifying questions and do not reconcile conflicts. They parse one surface and propagate whatever it says, so when a project's surfaces disagree about its own authors, version, identifier, or license, credit and provenance quietly follow whichever file a machine happened to read first. This poster will give RSEs a practical answer: a self-audit workflow that takes minutes per project. Using an open-source pipeline, we harvested the metadata surfaces a project exposes, normalized each to a canonical record, and scored every pairwise combination with a multi-level rubric (exact, minor, conflict, missing), with stratified hand-verification to measure the pipeline's precision. The pipeline's precision was measured by stratified hand-verification. The audit produced a gap map for the project, organized around the recurring failure modes the underlying study verified: missing metadata, stale archive records, paper-versus-software conflation, registry ghost metadata, identity ambiguity, and version-versus-concept DOI confusion. The workflow is grounded in evidence. We report results from a consistency audit of 117 open-source research software projects, 87 from HPC and quantum computing plus a registered 30-project baseline drawn from JOSS and pyOpenSci [8], showing which surfaces disagree most, on which fields, and what the failure modes look like in real projects. Of the 62 projects exposing at least two comparable surfaces, 52 (83.9%) contain at least one core-field conflict, and half of hand-verified conflicts trace to a surface describing the software's paper rather than the software itself. Attendees leave with three things: the workflow, a short checklist for making their own project's metadata agree with itself, and a concrete way to reason about a question this year's theme raises directly: what changes about our work when its primary readers include automated machines?
Source: arXiv (Cornell University)
Author: Shan, Pengyin
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203759813
Cites:
Cited by:
Related to: