Oncofusion

start / graph / concept-a-pi-dimension-can-be-saturated-in-the-corpus-and-absent-from-the-entry-named-for-it

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q127-pi-dimensions-at-the-centre

A dimension can be saturated in the corpus and empty in the entry named for it, and the reporter that measures coverage cannot tell the difference

also called the write-back gap · corpus coverage is not entry coverage

This project had written a great deal about how KMT2A behaves in normal cells. None of it was in the file called KMT2A. A count of how much the project knows would have looked fine; the page a reader actually opens was empty.

The same thing, for a scientist

Measured this cycle: of 16 sources load-bearing for cell type, cell state and epigenetics on KMT2A and MEN1, 7 were already cited in this corpus - one in 32 entries, one in 24 - and 0 of the 16 appeared in gene-kmt2a [E1]. The conclusion of the single most-cited of them, that MLL1's methyltransferase activity is dispensable for haematopoiesis and MLL-AF9 leukaemogenesis, is written in this graph in at least four places and was absent from the entry for the enzyme. This is a distinct failure mode from the ones already recorded here, and the existing PI-dimension reporter is structurally unable to see it, because it scores entries independently and a dimension can be 'covered' corpus-wide while the entry a reader opens carries nothing.

Go deeper
The shape of the failure, and why it is not one of the ones already written down

This corpus already records several coverage failures:

  • Rule 89 - a coverage metric counted a link near the thing instead of the thing, and six highest-degree entries scored 100% with no localization record.
  • Rule 90 - a computed set's identities were thrown away and only the count kept, so a number went stale invisibly.
  • The five-dimensions note - the reporter never printed epigenetics at all, so nobody filled it.

All three are defects in the measuring. This one is not. The reporter can be entirely correct, print epigenetics, and still miss this, because the corpus knowing something and the entry saying it are different facts and only the second is what a reader gets. gene-kmt2a was, until this cycle, an entry about domains, coordinates, PDB coverage, partner frequencies and DepMap - a structural and epidemiological entry for a gene whose defining activity is an enzymatic one.

Why the highest-degree entries are the most exposed

Two reinforcing reasons, both visible in this project's history.

  1. They look complete. Rule 89's worked case is exactly this: a perfect score on the most important rows was read as reassurance rather than as a symptom. gene-kmt2a has 229 edges and eleven evidence items; nothing about it invites a reader to check whether it is missing the most-cited paper about its own enzyme.
  2. Nobody is dispatched to them. Beats are dispatched to gaps, and a gap is named by something absent from the corpus. An entry-level gap inside a well-covered topic is invisible to that process by construction - which is how two of the three genes this project is named after ended a corpus of 2,396 entries with two of five PI dimensions empty.
What to do, stated so it can be checked
  1. When a beat establishes a conclusion about gene X, it writes that conclusion into gene-x as well as into its own new entry. Not a link - a field or an evidence item. This beat's amendments to gene-kmt2a and gene-men1 are the worked example.
  2. Before writing a new entry about a gene, run scripts/check-citation-reuse.py on your PMIDs and then check specifically whether the gene's OWN entry cites them. The script answers a corpus question; the entry question needs one extra grep and is the one that has been failing.
  3. A per-entry completeness check is worth building: for each of the top N entries by degree, what fraction of the most-cited papers on that entry's topic does the entry itself cite? This beat measured it for two entries by hand and got 0 of 16 and 1 of 16. Whether that generalises is unknown and is the obvious next measurement.
Facts and figures
failure modewrite-back gap: a conclusion is established and cited across many entries and never lands in the entry for the object it is about
proposed checkfor each of the top N entries by degree, the fraction of the most-cited corpus papers on that entry's topic that the entry itself cites. Unbuilt.
distinguished from
{
  "rule_89": "metric counted a proxy (a link near the thing) - a defect in the measuring",
  "rule_90": "identities discarded, count persisted - a defect in the recording",
  "five_dimensions_note": "reporter never printed epigenetics - a defect in the output",
  "this": "measurement, recording and output can all be correct; the entry is still empty"
}
worked case
{
  "entry": "gene-kmt2a",
  "before_this_cycle": "no cell_type, no cell_state, no epigenetic_readout; 11 evidence items covering domains, coordinates, PDB coverage, partner frequency, gnomAD and DepMap",
  "most_cited_missing_source": "PMID 24813891 - MLL1 methyltransferase activity dispensable for haematopoiesis and MLL-AF9 leukaemogenesis - cited in 32 entries of this corpus and in none of gene-kmt2a"
}
How this could bite us back
  • Any argument in this project of the form 'the corpus already covers X, so no beat is needed' is exposed to this. Rule 90 records one stale number suppressing work on the PI's first priority for seven cycles; a corpus-level coverage figure that hides empty entries would do the same thing and leave no stale number behind to find.
Still open
  1. How many of the top 50 entries by degree are in this state? Two out of two checked, which is a terrible sample and a suggestive one.
  2. Does the same gap exist in the other direction - entries citing papers whose conclusions contradict the entry's own summary? find-contradictions.py detects same-PMID disagreements between entries but not an entry disagreeing with a paper it does not cite.
Every source behind this entry (1)
  1. 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, 7 were already cited in this corpus with a maximum of 32 entries for one source and 24 for another, while 0 of the 16 appeared in gene-kmt2a and 1 of 16 in gene-men1, that one being a deliberate positive control - measured over the 2,396 node files outside the measuring beat's own folder on 2026-09-23.

    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 2026
    What this evidence cannot show

    n = 2 entries. This is an existence proof that the failure mode occurs on the two most important entries in the project, not an estimate of how common it is. The 16 sources were selected by the beat after searching, so the corpus-level fraction is not a random sample; the entry-level fraction is exact for those two files and those 16 sources. See ana-q127-the-pi-dimensions-were-absent-from-the-entries-not-from-the-corpus for the full per-PMID table and for the instrument error in the first run.

    Evidence that argues against this (1)
    • The opposite shape was measured one cycle earlier on a different gene: 6 of 7 primary sources for wild-type NUP98 biology were genuinely new to the graph. So the corpus has both failure modes, and a beat arriving at a gene cannot assume which one it is looking at without measuring. Q126 beat 2026
Connected to