Measured: the cell-type, cell-state and epigenetics literature for KMT2A and MEN1 was NOT missing from this corpus - 7 of 16 load-bearing papers were already cited, one of them 32 times, and 0 of 16 were in gene-kmt2a
also called the Q126 comparison for KMT2A and MEN1
We checked whether this project had simply never read the papers about how KMT2A and MEN1 behave in normal cells. It had read them, across dozens of entries, and had written none of them into the two files named for those genes.
The same thing, for a scientist
Sixteen sources judged load-bearing for cell type, cell state and epigenetics on KMT2A and MEN1 were scored against the 2,396 node files outside this beat's folder, with two instruments: cited-as-evidence (a citation.pmid anywhere in the entry) and named-anywhere (a PMID scan of the serialised entry). Result: 9 of 16 never cited and 8 of 16 never even named; 7 already cited, including PMID 24813891 in 32 entries and PMID 17908926 in 24. And 0 of 16 appeared in gene-kmt2a, 1 of 16 in gene-men1 - and that one, PMID 19228930, was included deliberately as a positive control for the instrument. This is the OPPOSITE shape to Q126's NUP98 finding and it implies a different remedy.
The numbers
| measure | value |
|---|---|
| PMIDs assessed | 16 |
| never cited as evidence anywhere in the corpus | 9 |
| never named anywhere in the corpus, prose included | 8 |
| already cited | 7 |
| cited in 10 or more entries | 2 |
| highest count (PMID 24813891, MLL1 SET activity dispensable) | 32 |
| second highest (PMID 17908926, bimodal cell-cycle degradation of MLL) | 24 |
present in gene-kmt2a before this cycle | 0 |
present in gene-men1 before this cycle | 1 (the positive control) |
Baseline: 2,396 node files, excluding this beat's own 12. Persisted with the entry ids per PMID in analyses/q127-the-pi-dimensions-at-the-centre/results.json.
Why this is not Q126's finding
Q126 measured wild-type NUP98 biology and found 6 of 7 primary sources new to the graph, despite NUP98 being named in 221 node files. The diagnosis there was that the corpus knew the fusion partner and not the protein, and the remedy was to go and read.
Here the corpus knows the protein. It knows that the MLL1 SET domain is dispensable, it knows the fusion escapes cell-cycle degradation, it knows menin-MLL inhibition drives human beta-cell replication through CDKN1A/1B/2C. Those conclusions are written, cited and argued about across dozens of entries. What was missing was the write-back into the two highest-degree entries in the project.
The remedy is correspondingly different: not a reading beat, but a rule about where a conclusion has to land.
Two instruments, one disagreement, and the disagreement is the interesting part
scripts/check-citation-reuse.py answers is this paper cited as evidence here. A plain PMID scan answers is it named anywhere here. They disagree on exactly one of the sixteen.
PMID 40016386 - MI-503 in osteosarcoma, one of only two pharmacological measurements of the menin-CDK-inhibitor axis that exist - is not cited anywhere, and is named in ana-the-menin-parp-experiment-survives-and-gains-a-rival-rationale. It appears there as a false positive of a bounded menin AND PARP query, correctly discarded because its 'PARP' is cleaved PARP, an apoptosis readout.
That entry did exactly the right thing for its own question. The consequence is that a paper reporting a p27 result this project needed sat inside a screened-out control list, read once for the wrong feature and never read again. This is the same failure the corpus already records as a control count is a list nobody opens - here the list was opened, for a different purpose, and the relevant fact was still missed.
The instrument error in this analysis, recorded because it is instructive
The first run of this measurement used scripts/check-citation-reuse.py after this beat had already written eight nodes. The script walks every node file, so it counted this beat's own citations: PMID 18371366 came back as cited 4 times when the correct pre-beat answer is 0, and PMID 24813891 rose from 32 to 33.
Every one of those four was a file written twenty minutes earlier by the agent running the check. The corrected run excludes data/graph/nodes/q127-the-pi-dimensions-at-the-centre/ and persists the baseline entry ids rather than only the counts, because anyone re-running the shipped script after this cycle will get the inflated numbers and nothing in the output will say so.
| script | analyses/q127-the-pi-dimensions-at-the-centre/ (results.json, bounded-queries.json, citation-reuse-raw.txt); the measurement walks data/graph/nodes excluding the beat's own folder |
|---|---|
| status | run |
| result summary | 9 of 16 never cited, 8 of 16 never named, 7 already cited (max 32 entries); 0 of 16 in gene-kmt2a, 1 of 16 in gene-men1 (positive control). |
| n pmids | 16 |
| n never cited | 9 |
| n never named | 8 |
| n already cited | 7 |
| max entries citing one source | 32 |
| n in gene kmt2a before | 0 |
| n in gene men1 before | 1 |
| baseline node files | 2396 |
| contrast with q126 | Q126: 6 of 7 primary sources for wild-type NUP98 biology new to the graph - a corpus-level gap. Q127: the corpus already held the key sources and the two central entries did not - an entry-level gap. |
| known instrument defect | scripts/check-citation-reuse.py counts the running beat's own files. Run it BEFORE writing, or exclude your folder. |
inputs
[ "16 PMIDs selected by this beat", "2,408 node files total, 2,396 baseline", "scripts/check-citation-reuse.py output retained as citation-reuse-raw.txt" ]
- If the graph's highest-degree entries can be missing the most-cited papers about their own biology, then entry degree is not a proxy for entry completeness - and this project has repeatedly used 'the top entries look complete' as a reason not to check them.
- How many other high-degree entries are in this state? The measurement here is two entries. A corpus-wide version - for each of the top 50 entries by degree, how many of the 10 most-cited papers on that entry's own topic does it cite - would be a one-evening script and would say whether this is a pattern or a pair.
- Should check-citation-reuse.py exclude the caller's own folder by default, or print the split? It would have made this beat's first-run error impossible and the change is a few lines.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of 16 sources load-bearing for cell type, cell state and epigenetics on KMT2A and MEN1, 9 were never cited as evidence and 8 were never named at all across the 2,396 node files outside this beat's folder; 7 were already cited, two of them in 24 and 32 entries; 0 appeared in gene-kmt2a and 1 in gene-men1, that one being the deliberate positive control.
Citation-reuse measurement over the oncofusion knowledge graph, two instruments, baseline excluding the measuring beat's own files · Q127 beat, oncofusion knowledge graph · analyses/q127-the-pi-dimensions-at-the-centre/results.json 2026What this evidence cannot show
The 16 PMIDs are NOT a random sample. They are the sources this beat selected after searching, so the new-fraction is a property of the selection and is not an estimate of corpus coverage - a different beat choosing different sources on the same two genes would get a different fraction. The entry-level result (0 of 16 in gene-kmt2a) is robust to that objection in a way the corpus-level result is not, because it is a statement about two specific files and every one of the 16 was checked against them. PMID 19228930 is in the set as a positive control and scored present in gene-men1, so the substring instrument is not silently failing.
Evidence that argues against this (1)
- Q126, one cycle earlier, measured the same question for NUP98 and got the opposite shape - 6 of 7 primary sources new to the graph despite NUP98 being named in 221 node files. Both results can be true; they say the corpus's blind spots are gene-specific rather than systematic, which weakens any general claim drawn from either. Q126 beat 2026
- related to →KMT2A / MLL1 — the gene that breaks
0 of 16 load-bearing sources were in this entry before cycle 89
1 of 16 load-bearing sources was in this entry before cycle 89, and it was the positive control
the measurement behind the rule