What changes a fact-verification verdict? Matched evidence and reader switching on FEVEROUS
This preprint examines when a fact-verification system gains more from retrieving evidence or switching readers. On 160 FEVEROUS development claims, a 30B reader gave 44 correct verdicts with a claim-matched supplement, 33 with the article lead alone and 23 with equal-length donor text. The matched-minus-donor gain was 13.1 percentage points (95% group bootstrap interval 7.5–19.1). Matched text also increased incorrect SUPPORTS or REFUTES verdicts. A 120B reader scored 51 on the lead and 84 on the matched evidence. A policy limited to switching 40 smaller-reader answers scored 64, against 63 for a fitted control. The record includes the manuscript, source and an offline reproducibility supplement. FEVEROUS annotations and third-party inputs retain their applicable upstream terms. Source attribution, asset identifiers and reconstruction instructions are included in the supplement. I set the research direction, approved the evaluation question and directed model selection. I used Codex and Claude to help design and run the experiments, analyze and check the results, and draft and review the manuscript.
Authors
- Aryan Rajesh Ganjoo (ORCID: https://orcid.org/0009-0006-2060-6343)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24417644
- Primary Topic
- Misinformation and Its Impacts
- Type
- preprint