Counted absences keyed on a complex subunit miss the papers that name only the complex - demonstrated on our own SSRP1 entry, which declared no KMT2A paper about FACT while this graph cites one about the index fusion in four other entries

analyses/subunit-keyed-absence-sweep compares a symbol-keyed query with a complex-phrase query for 43 subunits across 15 complexes this graph models. 26 of 43 under-reach their complex by three-fold or more, and 14 return zero on the symbol while the complex returns records.

That table should not be quoted, and the reason is the trap this project logged two days ago. Title-checking the complex hits shows heavy contamination by ordinary English: "Mediator" returns mediator and mediated - all five top titles are noise. "FACT" returns in fact, four of five. "SEC" and "NELF" are the same shape. This is the "SET-NUP214" finding again: 1,421 records whose first ten titles were quantum chemistry and botany.

So the sweep proves nothing by counting and one thing by reading. The ratios are unusable; the single title-checked hit is decisive. That asymmetry is the finding, and it is the fifth time in this project that a score produced candidates and only reading produced a result.

Where the complex phrase is unambiguous the ratios probably do mean something - P-TEFb returned five relevant titles out of five, and PRC1/Polycomb four of five - but those need the same reading before use.

It removes an absence and adds a positive, and neither touches the dependency data.

gene-ssrp1 and protein-fact-complex can no longer say that no KMT2A study concerns FACT. What exists is stronger than a mention: FACT is enriched at MLL-AF4 target enhancers, and losing it abolishes enhancer-promoter contact in the index fusion 2.

The DepMap picture is unchanged and still points the other way. SSRP1 is less required in KMT2A-rearranged lines than in other leukaemia lines (difference +0.265, permutation p 0.0104), at a pan-essentiality of -1.271 where both arms are dying. These are not in conflict. A factor can be required for a specific enhancer-promoter architecture and still be required no more than everywhere else for survival - and a knockout screen measures survival.

The two together are a better entry than either. A structural requirement with no differential survival cost is precisely the shape of a target that a viability screen cannot find, which is the argument the curaxin result was already making from the drug side 3.

Not re-run them mechanically. The sweep shows the automated version produces a contaminated table, and 172 entries is more reading than a cycle holds.

The rule that costs nothing is prospective and belongs in the brief: before writing a counted absence on a protein that belongs to a named complex, run the complex form too, and title-check it. That is one extra query at writing time against an unbounded re-reading cost later.

The retrospective queue should be ordered by stake, not by ratio. The absences worth re-reading first are the ones a conclusion rests on - gene-ssrp1 was one, and it was found because a beat happened to sweep past it. Candidates on the same pattern: NELF, DSIF, CAF-1, the HIRA/HUCA complex and Mediator, all of which returned zero on every subunit symbol tested.

And there is a second-order lesson. This graph cited PMID 37626123 in four entries while a fifth entry declared the literature empty. Nothing checks an entry's absence claim against the graph's own bibliography, and that check is cheap: a counted absence whose subject matches a paper the graph already cites should be flagged automatically.