Testing this corpus's own recorded refusals instead of its gap list overturned the first refusal tried, and the refusal had named the falsifier itself
also called re-test the refusals not the gaps · the HDAC6 refusal did not survive · a [tiab] bound is not a bound · search the reagent, not the disease
Last cycle showed this project's list of missing gene write-ups was empty - everything on it had already been decided against. So this cycle tested the decisions themselves. The very first one checked turned out to be wrong, and the evidence that refutes it had been published five years before the refusal was written.
The same thing, for a scientist
Cycle 142 established that **0 of 15** ids on this corpus's unwritten-gene worklist were gaps: 14 were deliberate refusals and one was not an interactor. Its conclusion was that **re-testing the refusals is a better lead than chasing gaps**, because a refusal carries a reason and a reason can go stale. This cycle tested that, and it worked immediately. 🔴 **The first refusal tested did not survive.** `ana-no-hdac-inhibitor-result-in-this-disease-attributes-to-an-enzyme` declined HDAC6 on the ground that `(MLL[tiab] OR KMT2A[tiab]) AND HDAC6[tiab]` returns **0 records**, and named its own falsifier: *"any HDAC6-selective agent tested in a KMT2A-rearranged model."* Replacing the disease terms with **drug names and cell-line names** returns **15 records**, of which **6 are on-beat** after reading titles. Among them: **PTG-0861**, HDAC6 IC50 **5.92 nM**, ~36x isozyme-selective, active against **MV4-11** - which is **KMT2A::AFF1**, this project's index fusion (PMID 32615502, 2020); and an artemisinin-derived selective HDAC6 inhibitor with **MV4-11 IC50 28 nM** and a **MOLM-13-Luc xenograft** (PMID 41701111). **None of the seven papers is cited anywhere in this corpus**, and no HDAC6-selective agent was named in any node file [E1]. The refusal was written against a literature that already existed. `gene-hdac6` is now written, with the counter-evidence that keeps it honest. 🔵 **The instrument defect was already classified here, which is why this is a method finding and not a lucky hit.** This corpus's own bounded-absence audit found **23 of 39** bounded absences were *ABSTRACT-LEVEL literature search*, explicitly recorded as **not a bound**, and named a defect class of *"entries that ask PubMed a question PubMed cannot answer."* The HDAC6 refusal is a textbook member of that class. **The generalisable rule: when the claim is "no agent has been tested in model X", query the agents and the model, not the disease** - a medicinal-chemistry paper reports `MV4-11 IC50 = 28 nM` in a potency table and has no reason to write *KMT2A* anywhere. 🟠 **Two of the three refusals tested survived, and saying so is the point.** **SIRT6**: the falsifier was *"an NAD+-dependent deacetylase result at a fusion target locus"*; the query returns 2 records, one the already-dismissed DPY30 paper and one a nasopharyngeal-carcinoma ferroptosis study (PMID 40081214) that reaches SIRT6 through HOXA9 but is not leukaemia and not a locus result. **Does not fire.** **KAT6B**: the falsifier was *"a lymphoid or infant result"*; the strongest hit is a 2025 *Cancer Discovery* paper combining KAT6/KAT7 inhibition with menin inhibitors in MLL leukaemia - **myeloid, and already cited in this corpus**, with `gene-kat7` and `target-phase2-kat7-dossier` already written. **Does not fire** [E2]. **1 of 3 refusals overturned** is the honest headline, and the two survivors are what make the first one credible.
What I nearly got wrong twice in the same hour
🟠 I almost reported KAT7 as a gap. The KAT6B search surfaced a 2025 Cancer Discovery paper reporting that KAT7 interacts with menin and the MLL complex, that KAT6/KAT7 inhibition evicts the MLL fusion from chromatin, and that it overcomes primary and acquired resistance to menin inhibitors - which bears directly on this project's central hypothesis. My instinct was that this must be missing. It is not: scripts/is-symbol-covered.py returns COVERED for KAT7, the corpus names it in 97 node files, PF-9363 appears in 4, and the paper's PMID is already cited. Running the tool first cost seconds and prevented cycle 142's error from recurring a fourth time.
🔴 And I misattributed an author while writing. A first draft of gene-hdac6 credited PMID 32615502 to "Tavares MT, et al." - a name carried over from a different paper in the same result set. The real first author is Gawel JM and the senior author Gunning PT; the volume was wrong too (201, not 200). Worse than the slip: the caveat I wrote claimed the author list came from the PubMed record, which it did not. Both were re-fetched and corrected, and the correction is recorded in the entry's own caveat.
🔵 The pattern is the same one this corpus keeps recording: every identifier I fetched before writing was right, and the one I typed while writing was wrong. An author list is an identifier.
🟠 Nine of fifteen search hits were discarded on reading, and two of the discards are instructive: SEM matched a duckweed paper and a catfish paper, because PubMed expands it to scanning electron microscopy and standard error of the mean; THP-1 matched four inflammation papers using it as a monocyte model rather than a leukaemia. A cell-line name is not a safe query token.
| refusals tested | 3 |
|---|---|
| refusals overturned | 1 |
| refusals survived | 2 |
| tiab query records | 0 |
| reagent query records | 15 |
| on beat after title reading | 6 |
| discarded on reading | 9 |
| near miss avoided | KAT7 looked like a major gap after the Cancer Discovery paper surfaced; is-symbol-covered.py returned COVERED, 97 node files name it, PF-9363 appears in 4, and the PMID is already cited |
| self correction | a first draft of gene-hdac6 attributed PMID 32615502 to the wrong first author, carried from an adjacent paper in the same result set, and its caveat falsely claimed the list came from the PubMed record; both fixed after re-fetching |
| new entry written | gene-hdac6 |
overturned
[ "HDAC6" ]
survived
[ "SIRT6", "KAT6B" ]
discard reasons — the numbers behind the chart
{
"SEM expanded to scanning electron microscopy / standard error": 2,
"THP-1 used as a monocyte inflammation model": 4,
"non-leukaemia or non-selective": 3
}papers newly cited by this cycle
[ "32615502", "41701111", "26443078", "39067437", "22464548", "40370107", "38148433", "25688158", "41249257", "29643474", "41484929", "40081214" ]
measurements
[
{
"quantity": "recorded refusals overturned when re-tested against their own falsifier",
"value": 1,
"unit": "refusals",
"subject": "HDAC6, declined on a [tiab] query returning 0 records",
"universe": "3 refusals tested this cycle (HDAC6, SIRT6, KAT6B); 2 survived"
},
{
"quantity": "PubMed records returned by a reagent-and-model query where the entry's [tiab] query returned zero",
"value": 15,
"unit": "records",
"subject": "HDAC6-selective agents against KMT2A-rearranged cell lines",
"universe": "6 on-beat after reading titles; 9 discarded, including 2 SEM false positives"
}
]- Re-run every bounded absence whose bound is an abstract-level literature search. The corpus's own audit counted 23 of 39 in that class, and this cycle shows the class is not merely theoretically weak - the first member tested was wrong.
- Decide whether a [tiab] query should be allowed to bound an absence at all. On this evidence it should be recorded as a search result, never as a bound, and entries that rely on one should carry a standing flag.
- Test an HDAC6-selective agent in RS4;11 or SEM. Both existing literatures stop one cell short of the index population, which is the single cheapest experiment this entry points at.
- Check whether this corpus's KAT7 coverage includes the menin-interaction and menin-resistance results specifically, rather than only citing the paper. A cited PMID is not a covered finding, as this project has recorded before.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
The HDAC6 bounded absence in ana-no-hdac-inhibitor-result-in-this-disease-attributes-to-an-enzyme is overturned by its own stated falsifier: a PubMed query over drug names and cell-line names returns 15 records where the entry's [tiab] query returned 0, and at least four HDAC6-selective inhibitor series have been tested against MV4-11, a KMT2A::AFF1 line, including PTG-0861 (HDAC6 IC50 5.92 nM, ~36-fold selective) and an artemisinin-derived lead (HDAC6 IC50 12 nM; MV4-11 IC50 28 nM, MOLM-13 IC50 89 nM, MOLM-13-Luc xenograft). None of these papers was cited in this corpus and no HDAC6-selective agent was named in any node file.
PTG-0861: A novel HDAC6-selective inhibitor as a therapeutic strategy in acute myeloid leukaemia · Gawel JM, Shouksmith AE, Raouf YS, Nawar N, Toutah K, Bukhari S, Manaswiyoungkul P, Olaoye OO, Israelian J, Radu TB, Cabral AD, Sina D, Sedighi A, de Araujo ED, Gunning PT · Eur J Med Chem 201:112411 2020What this evidence cannot show
All PMIDs and DOIs were fetched from PubMed on 2026-09-24 and their titles and abstracts read; full texts were not obtained. The overturn is of the entry's own falsifier clause, which asked only for an HDAC6-selective agent tested in a KMT2A-rearranged model - a low bar that says nothing about whether HDAC6 matters to the fusion. The parent entry's broader headline is weakened rather than refuted. MV4-11's KMT2A::AFF1 status is taken from this graph's own drug-chidamide entry rather than re-verified against a cell-line authority.
Evidence that argues against this (2)
- Overturning a falsifier is not the same as establishing biology. The evidence is antiproliferative potency in cell lines, the same compounds are comparably potent in HL60 and MM1S which carry no KMT2A rearrangement, and a dedicated isoform comparison in B-ALL (PMID 25688158) concluded HDAC1/2 rather than HDAC6. The refusal's conclusion may have been right for the wrong reason.
- One overturn out of three tested is a small sample, and HDAC6 may have been the easy case precisely because it is a heavily drugged target with a large medicinal-chemistry literature indexed under compound names. Refusals about poorly-drugged proteins would not be rescued by this query change, so the 1-of-3 rate should not be extrapolated to the corpus's other bounded absences.
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.journal-article
Two of the three refusals tested this cycle survived: the SIRT6 falsifier (an NAD+-dependent deacetylase result at a fusion target locus) returns 2 records, one previously dismissed and one a nasopharyngeal carcinoma ferroptosis study reaching SIRT6 via HOXA9 and UBR5 rather than a leukaemia locus result; and the KAT6B falsifier (a lymphoid or infant result) is not met by the strongest available hit, a myeloid MLL-fusion study of combined KAT6/KAT7 and menin inhibition which is already cited in this corpus and whose KAT7 subject already has both a gene entry and a phase-2 dossier.
Catalytic Inhibition of KAT6/KAT7 Enhances the Efficacy and Overcomes Primary and Acquired Resistance to Menin Inhibitors in MLL Leukemia · Gordon SJV, Perner F, MacPherson L, Fennell KA, Wenge DV, Bourgeois W, et al.; Armstrong SA, Dawson MA · Cancer Discov 15(10):2117-2138 2025What this evidence cannot show
Both survivals are abstract-level judgements from PubMed searches run 2026-09-24 and inherit exactly the weakness this entry criticises: an abstract-level query is not a bound, so 'does not fire' here means 'no firing evidence found by this query', not 'no such evidence exists'. The KAT6B falsifier asked specifically for a lymphoid or infant result and the KAT6/KAT7 paper is myeloid, but the paper does define a KAT6B contribution in MLL-fusion leukaemia, so a stricter reader could call this a partial fire. Whether the corpus's existing KAT7 coverage includes the menin-interaction result specifically was not checked beyond confirming the PMID is cited.
Evidence that argues against this (1)
- Scoring my own falsifier tests risks grading leniently on the ones that would cost work. The KAT6B case is the honest worry: the Cancer Discovery paper defines KAT6A, KAT6B and KAT7 contributions in MLL-fusion AML, and the original refusal rested on a DepMap rank rather than on absence of literature, so the refusal's stated reason and its stated falsifier are not about the same thing.
the HDAC6 clause is overturned by the falsifier the entry itself named
the entry written as a result of the overturn
acts on cycle 142s conclusion that refusals are a better lead than gaps
cited the tool as reliable; that still holds for a plain invocation
applies the cycle-143 agent-and-model instrument to eleven more absences; zero further overturns
the SEM collision that entry documented fired again one cycle later, because it was never added to the registry