Andrew Fribush

fig. 2.22.2: nine numbers and claims in four papers that are not in the paper they cite, printed against what the cited source says

fig. 2.22.2 · nine numbers and claims in four papers that are not in the paper they cite, printed against what the cited source says.
NINE NUMBERS AND CLAIMS IN FOUR PAPERS ARE NOT IN THE PAPER THEY CITEONE ROW IS ONE CITATION. ACCENT: WHAT THE PAPER PRINTS. INK: WHAT THE CITEDPAPER HAS. A ROW OPENS ON THE SOURCE.ELMO · TABLE 9: 76.5 EM, 84.3 F1HAS 71.3 EM SINGLE MODEL, 75.9 EM ENSEMBLECITED: R-NET (WANG ET AL. 2017)CAME FROM: A LATER, UNPUBLISHED LEADERBOARD ENTRYCITED SOURCE · V128 NOT SUPPORTED · LINE 1632“The single model achieves 71.3% on the evaluation metrics of exact match on the hidden testset, while the ensemble model further boosts the results to 75.9%.”

Each verdict quotes the cited source, and the pipeline checked every quoted span as a verbatim substring of the fetched text. Where the span the judge relied on is table cells, the row opens on the judge’s finding instead. Nilsson et al. 2008 is cited twice for bees (V014, V017). No verdict has yet been checked against a human.

source: replicatione docs/findings-2026-09-08.md lines 85-136 (section 4); source spans and findings verbatim from workspace/papers/<id>/report.md (ELMo lines 1321, 1538, 1632; insect biomass 886, 901, 1107; SQuAD 596, 721; subword units 1215, 1325) · citationChecks.data.ts

View in context: What 1,097 citation checks found in fourteen well-known papers → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.