Retrieval error in patent corpora: a repair-first pipeline and reporting items for the USPTO Open Data Portal transition

Patent landscapes are difficult to reproduce when retrieval failures, metadata repairs, and changing denominators are hidden. This article develops a repair-first pipeline in which deficits remain versioned negative results and each 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 during the transition from PatentsView to the authenticated USPTO Open Data Portal (ODP). The study separates ingest, canonical, attempted, and scored surfaces rather than mixing their denominators. Of 31,242 parseable-US records present at the 10 June 2026 repair run, ODP resolved 30,458 (97.49%); this is 20.05% of the later frozen canonical corpus. The repair surface is US-only by design: ODP carries file-wrapper authority for US documents, so records from other jurisdictions remain explicitly undated negative results in this iteration. A 10- to 11-digit pre-grant publication-number transform recovered 18,673 records, showing that 95.97% of first-pass failures were false absences created by query format. A reproducible title check across all returned records found 29,819 exact matches, 69 prefix-compatible matches, and 570 discordant retrievals (1.871%; Wilson 95% CI 1.725–2.030%). A stratified sample of 1,039 records, adjudicated under a pre-published protocol, found one identity failure among the discordant records (0.156% of the 639 examined) and bounds the identity error rate across the returned US records at 1.88% (95% confidence); the remaining discordances resolved to transcription or prosecution artifacts rather than different documents. Coverage is reported as a measured ramp rather than an asserted date boundary. A separate EPO–ODP assignee comparison on 26,644 US documents illustrates why cross-source agreement is a consistency check, not proof of fidelity. Drawing on defects exposed by the audit, the article proposes thirteen reporting items complementary to the Reporting Items for Patent Landscapes statement. The frozen attempted set, hashes, raw responses, scripts, and corrected artifacts are released for reproduction.