Only one query was broken, and it was dragging the whole figure down
also called narrowed recall · query ladder · 9.8% corrected
After one of our five literature searches turned out to be five times too pessimistic, we re-checked the other four. They were fine, and the overall figure we had published was being dragged down by the one broken search.
The same thing, for a scientist
The cycle-48 DOT1L correction left a worse problem than it fixed: if one of five queries was inflated five-fold, the pooled 9.8% and every per-beat comparison drawn from it were untrustworthy. Narrowing the other four the same way shows **they were not contaminated**. Adding a leukaemia restriction is a **no-op for all four** — each already contained a leukaemia term — and restricting further to KMT2A/MLL moves recall by a factor of **1.00, 1.01, 1.17 and 1.63**. Pooled over the four the figure is **15.0% unrestricted and 15.8% on the KMT2A-restricted slice**, against the 9.8% published across all five. **DOT1L was not one bad query among five; it was the only bad one, and it was dragging the aggregate down.**
The ladder
| query | T0 primary | T0 recall | T1 (+leukaemia) | T2 primary | T2 (+KMT2A) | T2/T0 |
|---|---|---|---|---|---|---|
| kmt2a-rearranged leukaemia | 833 | 17.0% | 17.0% | 833 | 17.0% | 1.00 |
| menin in leukaemia | 145 | 35.2% | 35.2% | 109 | 41.3% | 1.17 |
| infant acute lymphoblastic leukaemia | 175 | 16.6% | 16.6% | 100 | 27.0% | 1.63 |
| elongation-complex fusion partners | 436 | 15.4% | 15.4% | 399 | 15.5% | 1.01 |
| DOT1L/H3K79 (re-run in cycle 48) | 354 | 3.7% | 16.4% | 56 | 19.6% | 5.3 |
| pooled over the four | primary | cited | recall |
|---|---|---|---|
| T0 | 1,197 | 179 | 15.0% |
| T1 | 1,197 | 179 | 15.0% |
| T2 | 1,049 | 166 | 15.8% |
T1 being a no-op is the result. All four base queries already carry a leukaemia term, so there was nothing for a disease restriction to remove. DOT1L was the one query written without one — DOT1L[tiab] OR "H3K79"[tiab], against a gene studied in reprogramming, solid tumours and basic chromatin biology.
Infant ALL is the one with a real narrowing effect, and it is the informative one
At 1.63 the infant-ALL query is the only one of the four where restricting to KMT2A/MLL changes the answer much: 16.6% to 27.0%. That is not contamination in the DOT1L sense. Infant acute lymphoblastic leukaemia is a real field that includes non-KMT2A infant disease — hypodiploid, NUTM1-rearranged, germline-KMT2A — which this graph correctly does not read closely.
So the honest reading is a pair: the graph reads about a sixth of recent infant-ALL literature and about a quarter of the KMT2A-rearranged part of it. Both numbers are true and quoting either alone misleads in opposite directions.
What this means for the entry it corrects
ana-literature-recall has now been corrected twice: once for the DOT1L headline and once for the pooled figure that headline was distorting. The instrument is sound — T0 reproduces the parent's numbers, the parent's own counter-evidence predicted the DOT1L problem, and narrowing was the right response. What failed twice was publishing an aggregate before checking whether one of its components was broken.
The corrected picture is better than the published one and still not good: roughly 15% of the recent literature on this project's own queries is cited anywhere in 838 entries.
- The 65-laboratory absent list was computed on the contaminated pooled set. Recompute it on the narrowed set: some of those laboratories will be artefacts of the DOT1L query and will drop out, and the remainder is a better reading queue.
- Every recall figure here counts a paper as READ if any entry cites it. That counts a paper cited once in passing the same as one whose methods were read. No instrument here distinguishes them.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
Narrowing the four queries the DOT1L correction did not re-check shows none was contaminated: restricting to leukaemia terms is a no-op for all four because each already contains one, and further restricting to KMT2A/MLL changes recall by a factor of 1.00 (kmt2a-rearranged), 1.01 (elongation partners), 1.17 (menin) and 1.63 (infant ALL). Pooled over the four, primary-paper recall is 15.0% unrestricted and 15.8% on the KMT2A-restricted slice.
Narrowed literature recall: a three-tier query ladder over the four queries the DOT1L correction did not re-check · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run by the coordinator on 2026-09-22 with the same machinery, window (2020-2026), primary-only filter and definition of READ as the parent analysis and the DOT1L re-run. Per-query recall figures are 1-4 points above the parent's because the graph has grown by about a hundred entries since that run, so more of the same papers are now cited.
Evidence that argues against this (2)
- Narrowing raises recall by shrinking the denominator, so a high restricted figure is not evidence of thorough reading — it is evidence that what remains is what a KMT2A graph would cite. The pair has to be quoted together, which is precisely what the original headline failed to do. Oncofusion knowledge graph (internal analysis) 2026
- Four queries being clean does not validate the query SET. All five were written by the same person in one sitting to bound a field that has no canonical boundary, and a boundary drawn to be measurable is not the same as the field. The 65 absent laboratories the parent found included senior authors working on renal cell carcinoma, which is what a boundary drawn this way looks like from the inside. Oncofusion knowledge graph (internal analysis) 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
The published pooled figure of 9.8% across five queries was depressed by the one contaminated query: DOT1L/H3K79 contributed 354 primary papers of which 13 were cited, and narrowing it to KMT2A/MLL leaves 56 of which 11 are cited. Pooled over the four uncontaminated queries the figure is 15.0%.
Narrowed literature recall: a three-tier query ladder over the four queries the DOT1L correction did not re-check · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run by the coordinator on 2026-09-22 with the same machinery, window (2020-2026), primary-only filter and definition of READ as the parent analysis and the DOT1L re-run. Per-query recall figures are 1-4 points above the parent's because the graph has grown by about a hundred entries since that run, so more of the same papers are now cited.
Evidence that argues against this (2)
- 15.0% and 9.8% are not directly comparable and the entry should not pretend otherwise: the first pools four queries and the second five, and the graph has grown by about a hundred entries between the two runs, which lifts every recall figure by a point or more. The correction's direction is solid; its magnitude mixes three causes. Oncofusion knowledge graph (internal analysis) 2026
- Being corrected upward is the more comfortable direction and deserves the same scepticism as a correction downward would get. Roughly 15% of the recent literature on this project's own chosen queries is cited anywhere in 838 entries, which is a low number however it is decomposed. Oncofusion knowledge graph (internal analysis) 2026