The referral-denominator audit: 97 frequency claims, 63 qualified, and one real defect where the ledger predicted seven
also called ledger item N · recombinome referral bias · denominator audit
Most of what this graph knows about which gene partners KMT2A is broken with comes from one reference laboratory that receives samples sent to it from around the world, so its numbers describe who gets referred rather than who gets ill. The project's own to-do list said a warning about this had been copy-pasted into at least seven entries and that later corrections had reached only one of them. Checking it properly, that is not what happened: the warning sits verbatim in two entries, and of the ninety-seven places where the graph quotes a number from that series, sixty-three already carry the warning and exactly one was missing it in a way that mattered.
The same thing, for a scientist
Ledger item N asserted the recombinome referral-bias note was copy-pasted verbatim across at least seven `fusion-kmt2a-*` entries with corrections reaching one sibling and not the rest. **It does not reproduce.** The verbatim note appears in two entries with identical text. Scoped correctly - to evidence items that state a frequency or share sourced to the series, which is where the defect can actually live - the graph has **97 such claims, 63 carrying a referral/ascertainment qualifier and 34 not**. Reading the 34: most carry a different and better-targeted caveat, and **one** was a genuinely unqualified population claim. Fixed.
The ledger item did not reproduce, and the scoping is why
Entry-level matching is the wrong instrument and gives a meaningless number. A first pass asking 'does this entry mention the recombinome and lack the word referral' returned 195 entries, 86 of them 'missing' the qualifier. That number is worthless: most of those entries cite the series for a breakpoint or partner fact that carries no frequency claim at all, and a claim with no denominator needs no denominator caveat.
The defect can only live at evidence-item granularity, in an item that (a) states a frequency or share, (b) sources it to the series, and (c) does not record in its own caveats that the denominator is referrals. Scored that way:
| population | n |
|---|---|
| frequency claims sourced to the series | 97 |
| carrying a referral/ascertainment qualifier | 63 |
| missing it | 34 |
| of those, genuinely unqualified population claims | 1 |
This is the third time in two cycles that one of my own detectors over-flagged and reading separated defect from design.
What the 34 actually were
Reading them rather than counting them:
- Most are series-descriptive, not population claims. '47 of 3,401 genomically characterised rearrangements' is a statement about the cohort's own composition. Referral bias does not distort a description of the referral series itself.
- Several carry a different and better caveat than the one I searched for.
breakpoint-kmt2a-minor-bcrsays the BCR1:BCR2 ratio is method-dependent and that 'the true prevalence of BCR2 is uncertain and probably higher than 1.4%' - which addresses the denominator problem more precisely than the word 'referral' would have, and my regex could not see it.hyp-the-yeats-asymmetry-is-breakage-geography-not-selectionrecords that therapy-induced status is 'as submitted, not adjudicated', which is the ascertainment caveat that matters for that particular claim. - One was a real gap.
epi-kmt2a-mllt1-disease-and-age-distribution-460-casescarried 'MLLT1 accounts for about 5% of all MLL rearrangements' - a population claim quoted from a 2007 letter - with a caveat about the paper's partial unreadability and nothing about the denominator. A denominator note has been added.
The lesson repeats: a checker that searches for one phrasing of a caveat will score a better-worded caveat as absent. The same failure shape as scoring a claim sentence without reading its sibling caveats field.
- Anyone using this graph's partner-frequency numbers to size a trial arm or a diagnostic panel is using shares of a referral series. They over-represent cytogenetically cryptic and therapy-related disease by construction, and the graph's own minor-BCR entry states the true prevalence is probably higher than the series figure.
- Do not read '1 defect in 97' as licence to trust the frequency claims. It means the caveats are present, not that the underlying numbers are population frequencies - they are not, and no caveat makes them so.
- Does a claim that implies a population frequency without stating a number escape this audit?
- Should the referral-denominator caveat be a single concept entry that others link to, rather than prose repeated in 63 places? The repetition is currently a feature - each is specialised to its claim - but it is 63 places to correct if the series is ever re-characterised.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
The graph contains 97 evidence items stating a frequency or share sourced to the KMT2A recombinome reference-laboratory series; 63 record a referral or ascertainment qualifier and 34 do not, and of those 34 exactly one was an unqualified claim generalising to a population rather than a description of the series' own composition.
referral-denominator-audit: evidence-item level scoring of frequency claims sourced to the KMT2A recombinome series · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 against 1,466 entries. **The classification into 'population claim' versus 'series-descriptive' is regex-assisted and then read by hand**, and the boundary is genuinely fuzzy - 'MLLT3 is the commonest partner in therapy-induced KMT2A leukaemia' is both. **The qualifier test looks for a vocabulary, not a meaning**, and demonstrably scored at least two well-caveated items as unqualified because they made the same point in different words. So 63 is an UNDERCOUNT of the graph's real practice and 34 an overcount of the deficit. Moderate for that reason.
Evidence that argues against this (1)
- A single unqualified population claim out of 97 is a good result that could still understate the problem: the audit only sees claims whose text names a frequency, and a claim that implies a population frequency without stating a number - 'the commonest partner' with no count - may slip through the frequency regex entirely. Oncofusion knowledge graph (internal analysis) 2026
THE ONE GENUINE DEFECT: its '~5% of all MLL rearrangements' was a population claim quoted from a referral series with no denominator note. Corrected 2026-09-20.
SCORED AS MISSING THE QUALIFIER AND IS NOT: it says the true prevalence of BCR2 is uncertain and probably higher than 1.4%, which addresses the denominator better than the word the checker searched for.
- related to →Rolf Marschalek & Claus Meyer (DCAL, Goethe-University Frankfurt) — the world's KMT2A breakpoint dataset
THE SERIES' SOURCE: the concentration of this graph's partner-frequency evidence in one reference laboratory is what makes the denominator question load-bearing.
SAME FAILURE SHAPE: a checker that reads one phrasing of a caveat scores a better-worded caveat as absent, exactly as scoring a claim without its sibling caveats field did.