News · 20 October 2026
The patent-corpus reproducibility article is out in World Patent Information
“Retrieval error in patent corpora: a repair-first pipeline and reporting items for the USPTO Open Data Portal transition” is published in World Patent Information, volume 87, article 102511, doi:10.1016/j.wpi.2026.102511, in the journal's special issue on large language models, generative AI and agentic AI for intellectual-property analytics and strategy. The publisher's share link gives free access until 9 December 2026: authors.elsevier.com/a/1nzmkdA7TtalW.
Patent landscapes are hard to reproduce when retrieval failures, metadata repairs and changing denominators stay hidden. The article develops a repair-first pipeline in which a deficit remains a versioned negative result and every later repair records its source, status, identity check and retrieval time. The worked case is a 151,890-record patent evidence base for AI trust infrastructure, assembled while PatentsView gave way to the authenticated USPTO Open Data Portal. Ingest, canonical, attempted and scored surfaces are kept apart rather than mixed into one denominator.
Two findings carry the method. A 10- to 11-digit transform of pre-grant publication numbers recovered 18,673 records: 95.97% of the first-pass failures were false absences created by query format, not missing documents. A reproducible title check over all returned records, with a stratified sample of 1,039 adjudicated under a pre-published protocol, bounds the identity error rate across the returned US records at 1.88% (95% confidence). From the defects the audit exposed, the article proposes 13 reporting items for patent-information studies.
Data and code are on Zenodo, 10.5281/zenodo.21497997.