Working the queue deferred for four cycles: the 40 are three different things, and one absence claim the graph had already refuted ten entries away
also called stale absence recheck · absence queue triage
We keep a list of places where the graph says nobody has done an experiment, and we had not re-checked them for months. Checking found that most were fine, that many were not really claims at all, and that one had been answered by a paper the graph already knew about elsewhere.
The same thing, for a scientist
`check-stale-corrections.py` has reported 81 un-struck absence assertions, 40 of them before their correction, and the queue was deferred for four cycles. Reading it: the 40 are **section headings, the corrected text itself, and genuine absence claims** in roughly equal measure, so the reported number overstates the defect. Fourteen genuine, query-settleable claims were re-tested with query translations recorded. **One is genuinely stale**: `mechanism-menin-jund-transcriptional-repression` said a clinical menin inhibitor had not been tested against menin-JunD, and PMID 42371796 reports exactly that - **which the graph already cited in ten other entries**.
The queue is three different things, and only one deserves the name
| class | example | worth machine time? |
|---|---|---|
| section headings and titles | "Why this had never been tested", "What has never been done" | no - descriptions of history, not live claims |
| the corrected text itself | localization-endogenous-mll-af4-punctate-negative is flagged on the sentence that does the correcting | no - the checker is firing on its own fix |
| genuine absence claims | "nobody has published on METTL14 in KMT2A-rearranged disease" | yes |
ana-pattern-audit is flagged on a sentence that exists purely as a test case for this class of checker - the checker's self-referential exclusion list covers absence, framing-audit and five others but not pattern-audit.
So '40 un-struck assertions' is not 40 defects. That matters for how the number is read in the ledger: it is a reading queue, and reading it is most of the work.
Fourteen re-tested, one stale, and the one that matters is a resistance mechanism
Fourteen genuine absence claims that a PubMed query could settle were re-run, deliberately with queries broader than the claims so a true absence has to survive a generous test, and with querytranslation recorded for each because this project's documented route to a false zero is a parser-rewritten query.
Thirteen returned hits; twelve of those are off-target on reading - the Fanconi query's two hits are a murine Trem1 paper and a Shwachman-Diamond expression study, neither an FA gene tested in a KMT2A-rearranged cell, so that absence holds.
The exception is JUND. mechanism-menin-jund-transcriptional-repression said the pharmacology was missing: "it has not been reported for revumenib, ziftomenib or bleximenib". PMID 42371796 (Blood, 2026) reports it, and its own abstract confirms the claim was true when written - "the impact of menin-MLL inhibitors on JUND function has remained unexplored".
The finding inverts the entry's expectation: menin-MLL inhibitors do displace JunD, and displacement increases JUND chromatin occupancy, contributing to therapy resistance. A JNK-inhibitor combination is the proposed countermeasure. Recorded in that entry with bite_back.
The process failure is the point. The graph already cited PMID 42371796 in ten other entries, including gene-jund. Standing rule 45 says to propagate a correction by grepping every entry that cites the same PMID - but this entry does not cite it, it asserts the absence the paper fills. A citation-keyed sweep is structurally blind to that direction, and the absence queue is the only instrument that covers it. It had been deferred for four cycles.
- Add `pattern-audit` and any other self-referential id prefix to the exclusion list in `check-stale-corrections.py`, and teach it to skip `title` and `heading` fields - that alone would cut the reported 40 substantially without hiding a single live claim.
- Run this query sweep every cycle rather than accumulating a reading backlog. It took minutes and found the one thing four cycles of deferral had missed.
- The remaining 26 assertions before their correction, and the 41 after it, have not been re-tested. The ones that are not query-settleable need a different instrument or a human.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Of the 40 un-struck absence assertions sitting before their correction, a large share are section headings, titles, or the corrected sentence itself rather than live claims. Fourteen genuine, query-settleable claims were re-tested and thirteen returned hits, of which twelve are off-target on reading.
Re-testing the absence claims that sit before their correction against PubMed today, with query translations recorded · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
The three-way classification is this coordinator's reading, not a measurement, and the boundary between 'heading describing history' and 'live claim' is a judgement. Fourteen of forty were re-tested - the rest were either not query-settleable or judged not to be claims - so this is a worked sample, not a cleared queue. Only one query was run per claim; a second phrasing could change any of them.
Evidence that argues against this (1)
- Calling twelve of thirteen hit-sets 'off-target' is exactly the judgement an author has an interest in making, since each one that is on-target is a claim that has to be rewritten. The reading was done on titles, not full texts, and a title can hide a relevant experiment in a supplementary figure - this project has already found one case where a key result lived only in supplementary methods. Oncofusion knowledge graph (internal analysis) 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
The absence claim in `mechanism-menin-jund-transcriptional-repression` is stale. PMID 42371796 reports that menin-MLL inhibitors disrupt menin-JUND interactions in MLLr leukaemic cells and that the resulting increase in JUND chromatin occupancy contributes to therapy resistance. The graph already cited that paper in ten other entries without the correction reaching this one.
Menin-MLL inhibitors enhance JUND activity in MLLr leukemic cells contributing to tumorigenesis and therapy resistance · Ozyerli-Goknar E, Nizamuddin S, Fetsch V, et al. · Blood 2026What this evidence cannot show
Read from the abstract; the paper is online ahead of print and the full text was not retrieved. The count of ten other citing entries is a grep of the graph on 2026-09-20 and includes counter-evidence citations, so not all ten assert the finding.
Evidence that argues against this (1)
- One stale claim in fourteen is a low yield for a queue that has been carried for four cycles, and it was found by a query rather than by reading - which argues the queue should be **run as a query sweep every cycle** and read only where the sweep flags something, rather than treated as a reading backlog that never gets read. Oncofusion knowledge graph (internal tooling) 2026
its absence claim is stale: PMID 42371796 reports the menin-inhibitor experiment it said had not been done
one of ten entries already citing the paper the absence claim overlooked
third instance of a checker reporting a number nobody had read to ground
its test-case sentence is flagged by the stale-correction checker, which lacks it in the self-referential exclusion list
the FA absence claim re-tested and holds: both hits are off-target