Paper error pipeline: exported candidate findings (2026-10-23 22:18 UTC) adjudicated_majors.csv 5726 errors in 4482 papers (9400 top-ranked scale-run candidates checked). Each passed: a cheap full-paper screen (Claude Haiku 5.5 low or GPT-6 Luna low), Haiku 5.5 (high) and GPT-6 Luna (high) verification, then independent Opus 5.5 (high; Opus 5 where 5.5 refused) and GPT-6.1 Sol (high) checks that both called it a real major error. Quality checks on samples of this list: an adversarial defense (Sol) and judge (Opus) upheld 33 of 40 random rows as major and 30 of 40 of the lowest-ranked rows, with none judged a non-error; an independent agent audit reproduced 20 of 20 random rows as real errors and judged 13 of 20 major and uncorrected (20/20 and 18/20 for the highest-confidence rows). prereg_majors.csv 280 trial reports whose undisclosed discrepancy with the current ClinicalTrials.gov record was confirmed major by Opus after GPT-6 Luna and Claude Haiku both flagged it. Registry records can change after publication, so check the registry history before acting on a row. published_notice: a correction, retraction or expression-of-concern notice found for the paper (PMC, Crossref, Europe PMC). notice_fixes_this_error: "yes" if Claude Haiku 5.5 and GPT-6 Luna both judged, from the notice text, that it fixes the error in this row (treat those rows as already corrected); "no" if it concerns something else. pmc_copy: "author manuscript" means PubMed Central holds the authors' accepted manuscript (NIHMS), not the journal's final version. Errors are sometimes fixed during copy-editing: in our version-of-record checks, 5 of 18 such errors were absent from the publisher's version. Check the publisher's version before acting on these rows. all_experiments.csv one row per benchmark x experiment (screeners, unions, screen + verifier, multi-stage pipelines), scored with LLM-assisted matching. label/kind: readable name / screener, union, verified or pipeline. n_papers: papers scored (seed: 480 = 240 with one injected error + 240 unmodified controls; 240 for Opus, Sol and Sonnet, which screened the papers with an injected error only). failed_screen_rows: papers whose base screener output could not be used (mostly cut off at the output-token cap); they count as no flags, and unions and cascades built on that screener inherit them. n_known_errors: known errors in the benchmark; found: those matched by any output flag; recall = found / n_known_errors, with recall_lo/recall_hi its 95% case-bootstrap interval. n_major_errors: known major errors; major_held_major: those the output reports as major (the measure the report's single frontier pools over the four benchmarks). output_major_flags / major_flags_matching: denominator / numerator of precision_lower_bound (major flags on papers with a known error that match one; the papers hold other errors too, so true precision is higher). cost_per_paper / cost_total: LLM spend of all stages at current list prices. flags_per_paper: output flags per paper. major_false_alarms_per_clean_paper: major flags per seed control paper, an upper bound on false alarms (the unmodified originals can contain real errors); empty outside seed and where controls were not run. A flag that both screeners of a union raised is verified once, as in the scale run. These are model-verified candidates, not errors confirmed with authors or journals. Wording is factual and makes no claim about intent. Verify against the paper before contacting anyone.