A claim-to-paper overlap score measures where a citation's support lives, not whether it exists - so its zero is a pointer to a supplement, not a verdict against the citation
also called low overlap is not a misattachment · abstracts-only instruments and supplemental tables · a resource citation has no abstract to match
If you check a citation by comparing the words of a claim against the words of the paper's summary, a score of zero does not mean the citation is wrong. It usually means the evidence is in a table at the back of the paper, or that the thing being cited is a data deposit with no summary at all.
The same thing, for a scientist
An automated citation-support check that compares a claim's words against a paper's **abstract** produces three classes of low score, and **none of them is a misattachment**: 1. **The support is in a supplement.** A claim read from a supplemental table cannot overlap an abstract that does not mention the table's contents. Scored in this corpus: a claim about NLRC5 and CIITA group assignments scored **0.000** against the correct paper, whose abstract carries `menin` and `H2AK119ub` but none of `NLRC5`, `CIITA` or `MHC`. The entry's own caveat already said *"Read directly from supplemental Table 4."* 2. **The citation is a resource, not a paper.** A **PDB deposit** has a title and no abstract. For a claim about crystal packing in a specific structure, the deposit is the correct and only citation, and it will always score near zero. 3. **The claim is about the project's own method.** *"A PubMed query on 2026-09-20 combining revumenib, SNDX-5613 ... returned N hits"* is a claim about a search, not about any paper. It has no correct overlap value. 🔴 **The asymmetry that matters:** a **high** score is weak evidence of support (a menin paper is topically consistent with almost any menin claim), and a **low** score is weak evidence against it. So the instrument is nearly uninformative in both directions **as a verdict** - and genuinely useful as a **ranking** that tells a reader which pairs to open first. Its output is a worklist, not a rate. **Design consequence, stated as a rule:** an instrument of this kind must (a) refuse to issue a pass/fail verdict, (b) emit the ranked tail as its product, and (c) pre-declare the benign causes of a low score **before** the scores are seen, so that a zero is read as a pointer rather than a finding. The corpus's own run did all three, which is why its single 0.000 was correctly adjudicated as a supplement rather than filed as a defect.
The failure this does NOT cover
The dangerous citation error in this project is a working identifier pointing at a paper in the right field that does not support the specific number or direction claimed. Both documented historical failures here were of that kind, and both were caught by a human reading, not by any instrument.
A topical-overlap score is blind to it by construction: the wrong paper in the right field scores high. So this instrument's clean result must never be quoted as evidence that the corpus's citations support their claims - only that they are not attached to papers about unrelated subjects.
| pairs read | 5 |
|---|---|
| misattachments found | 0 |
| instrument blind spot | a wrong paper in the right field scores HIGH; both documented historical failures in this project were of that kind |
| design rule | emit a ranked worklist, refuse a pass/fail rate, and pre-declare the benign causes of a low score before seeing the scores |
low score causes observed
[ "support in a supplemental table", "citation is a data resource with no abstract (PDB deposit)", "claim is about the project's own query or method, not about the paper" ]
- Is there a fourth benign cause of a near-zero score that five read pairs did not surface? The 136 unread pairs are where it would be.
- Can the instrument be given a labelled positive? Neither documented wrong-paper failure falls in the placeholder-author subset, so its sensitivity is untested.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.
In a pre-registered run over the corpus's 34 least-verified citation identifiers, the single lowest-scoring pair (overlap 0.000) was a correctly attached citation whose support lies in a supplemental table, and the next three lowest were a PDB deposit cited for a crystal-packing claim, which has no abstract to match against.
What this evidence cannot show
Run 2026-09-24 (cycle 132). Five pairs of 141 were read by hand; the classification of low-score causes is drawn from those five and is not exhaustive. A fourth benign cause almost certainly exists that this sample did not surface.
Evidence that argues against this (1)
- It is possible for a claim to score low AND be misattached, and this run cannot distinguish those cases without reading - which is exactly why the instrument emits a worklist rather than a verdict. The concept therefore argues for reading the tail, not for dismissing it.
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.
The paper behind the 0.000-scoring pair is Zhou et al., 'Epigenetic regulation of noncanonical menin targets modulates menin inhibitor response in acute myeloid leukemia', Blood 2024, whose abstract describes noncanonical menin targets bivalently co-occupied by menin and H2AK119ub and the role of PRC1.1 subunits, and does not mention NLRC5, CIITA or MHC.
What this evidence cannot show
Abstract fetched 2026-09-24 from PubMed and read in full; the absence of NLRC5/CIITA/MHC was confirmed by string search of the fetched text rather than by impression. The supplemental table itself was not re-read in this cycle - the adjudication rests on the abstract's contents plus the cited entry's own provenance note.
Evidence that argues against this (1)
- Because the supplemental table was not re-opened here, this evidence item establishes only that the low score is EXPLAINED by the support being outside the abstract. It does not independently re-verify that supplemental Table 4 says what the citing entry reports; that remains on the original beat's reading and a separate beat's confirmation.
the general form of this run: a near-zero overlap locates the support rather than denying it
the resource-citation case: three pairs at 0.019-0.048 because a PDB deposit has no abstract to match
confirms the concept and adds a fourth cause: a claim that asserts an ABSENCE is scored backwards
quantifies the concept: an arbitrary same-field paper already scores 0.365 against these claims