[gabrielspadon]

Summary

The identity check treated the published SugarCrepe++ record (10.52202/079017-0571, arXiv 2406.11171) and the VISLA Benchmark preprint (10.48550/arxiv.2404.16365) as one work. They matched at a title score of 0.713. During the refresh replay this deleted VISLA from Milios, Oore, Sajjad and a2i2.

Root cause

_disk_identity rejected a preprint/published pair only when both records carried arXiv ids that differed (citeforge/identity.py, the preprint-pair branch). merge_utils strips eprint and archiveprefix from every record that has a published DOI, so a published record on disk never carries an arXiv id and the guard could not fire. The pair then fell through to SIM_PREPRINT_TITLE_THRESHOLD (0.55), which exists for retitled preprints. _enrichment_identity had the same gap in its composite gate: a Crossref candidate carries no arXiv id either, and VISLA vs SugarCrepe++ scored COMPOSITE True.

Fix

If only one record carries an arXiv id, that now counts as unknown identity rather than agreement. Unless both records carry the same arXiv id, the disk preprint-pair, composite and relaxed-preprint paths require SIM_FILE_DUPLICATE_THRESHOLD (0.95), and the enrichment composite gate stays closed. Identical-title twins, DOI matches, DOI versions, equal arXiv ids, external ids and truncated titles are unchanged.

Effect on the committed corpus

I evaluated every same-directory pair under DISK_SURVIVOR before and after the change. Five pairs flip from "same work" to "distinct":

• VISLA vs SugarCrepe++ in Milios, Oore and Sajjad. These are distinct works and are now kept.
• Oore, Lowe2023-ZeroShot.bib (2023 workshop paper) vs Lowe2024-EmpiricalStudy.bib (arXiv 2406.02465, TMLR 2026). These are distinct titles and are now kept.
• Reilly, Porcino2022-DesignXR.bib (arXiv 2208.14842, abbreviated title) vs Porcino2022-DesignExtended.bib (10.1007/978-3-031-20212-4_9). This is likely the same paper, but neither an arXiv DOI nor a Crossref relation links them, so the preprint is kept rather than removed as superseded.

Verification

• pytest -m 'not live' --cov=citeforge --cov=main --cov-fail-under=68 on Python 3.12, macOS: 2080 passed, 2 skipped (the documented cutover contracts in tests/test_workflow_contracts.py), 7 deselected, coverage 84.13%.
• ruff check citeforge/ tests/ main.py and mypy citeforge/ main.py are clean.
• test_published_record_never_absorbs_a_distinct_arxiv_preprint uses the committed VISLA and SugarCrepe++ records. Both directions fail on main: the incoming published record replaces VISLA, and an incoming VISLA is refused. Both pass with this change.
• test_enrichment_never_adopts_a_distinct_work_as_the_published_version fails on main with COMPOSITE True and passes with this change.

The refresh replay on main plus #60 and this change is reported in a comment below.

[gabrielspadon]

Refresh replay on main plus #60 and this change. This drove the real run_all over a copy of the committed corpus, with live Crossref, arXiv, OpenAlex and DOI enrichment of shared-DOI records. Scholar and Gemini were not used. finalize_run followed the replay.

• 9 authors were forced: Haque, Oore and Trappenberg, then Milios, Sajjad, Rudzicz, Zeh, Gagie and He. 0 DOI conflicts remained, and finalization completed with 3690 files and 1105 in a2i2.
• VISLA (10.48550/arxiv.2404.16365) is still present in Milios, Oore, Sajjad and a2i2.
• Per author against the committed corpus, 1 of 58 directories changed. Orji went from 267 to 266 files. D2025-DevelopingInteractive.bib (medRxiv 10.1101/2025.10.07.25337102) was removed as superseded by Etienne2025-DevelopingInteractive.bib (JMIR Human Factors 10.2196/86283). It is the same author list, retitled, and was matched at a title score of 0.773. This is not an arXiv pair, so this change leaves it unchanged, and the previous rule removed it too. Crossref carries no relation between the two DOIs.