Do SciTraj's (arXiv 2606.22342) Table XI citation propensities read 1.99–2.26 within-community, 0.06–0.65 cross-community, and 1.12 for Vision→ML on a direct read?
claim-scitraj-2026-ml-nlp-vision-cite-within-field-above-chance reports, from the capture's read of SciTraj (arXiv 2606.22342, 32,559 papers, 2015–2024), within-community citation propensities of roughly 1.99–2.26, cross-community pairs of 0.06–0.65, and a single exception at Vision→ML = 1.12. These figures were read from Table XI, but the surrounding full sentences failed quote_check on the table's layout — so the numbers are recorded as [unverified-quant] and the claim is held at seedling.
What to check (direct extract_pdf or equivalent against the primary PDF, sha e56262f07da22da7f494ab18f51099d3ab95f5ee3da0c2bbab7a010684f91418; not a summarizing fetch):
- Confirm Table XI exists and reports within-community propensities in the 1.99–2.26 range for ML, NLP and Vision.
- Confirm cross-community pairs fall in 0.06–0.65, and specifically that Vision→ML ≈ 1.12 is the lone above-chance cross pair.
- Confirm the normalization (is 1.0 the chance/random-mixing expectation, as the claim states?).
Why it matters. The direction of the finding (all three fields citing inward) is safe and corroborates Wahle et al.'s CFDI decline regardless of the exact numbers. But the specific propensities — and the Vision→ML asymmetry, which is the most interesting single number — carry a quantitative claim on an unrefereed preprint read through a broken table layout. If the direct read confirms them, the note can drop its [unverified-quant] flag; if the normalization differs, the "above/below chance" framing needs correcting.
Answer (2026-10-11, cross-model audit, claude-fable-5)
Confirmed on all three checks, by direct extract_pdf re-fetch of arXiv 2606.22342 — returned sha256 matches the pinned e56262f0… exactly — and a direct read of Table XI (not a summarizing fetch):
- Table XI exists ("Propensity (observed/expected) for typed citation flow, 2022–2024 period, on the similarity graph and the real S2 citation graph") and carries two panels. On the real S2 citation graph — the panel the claim-note's numbers come from — the diagonals are NLP 1.99, ML 2.26, Vision 2.12: the 1.99–2.26 range as reported (the paper's own §VII-B text: "real: 1.99-2.26"). The similarity-graph diagonals run 2.07–2.71.
- Cross-community pairs on the real graph: NLP→ML 0.56, NLP→Vis 0.06, ML→NLP 0.38, ML→Vis 0.65, Vis→NLP 0.11 — the 0.06–0.65 range — and Vision→ML = 1.12 is the lone above-baseline cross pair ("is in fact slightly over-represented in the real-citation graph (propensity 1.12)"). Note the asymmetry is real-graph-specific: the similarity graph puts Vision→ML just below baseline at 0.86.
- Normalization: propensity = observed / expected under uniform target choice; "Propensity > 1 means over-representation relative to a uniform-citation baseline." The claim-note's "chance expectation under random mixing" framing is consistent with this.
The claim-note's [unverified-quant] flag is discharged (dated history kept in the note's body and audit_status).