H3C1 stays rejected, and now with the two numbers the registry row lacked: its accession carries TEN gene names, and removing every substrate and ribosome symbol moves interactome coverage 43.8% -> 46.1%

The dispatch brief for this beat wrote: "H3C1 is histone H3.1 - the SUBSTRATE, not a conventional partner. There is a real question whether the histone belongs in an interactome census at all... Decide that explicitly and argue it; it is the most interesting judgement in this beat."

It was decided on 2026-09-20, two days before this beat was dispatched, and the row in data/symbol-collisions.json contains substantially the argument the brief asked for 1:

class: substrate_not_a_private_partner; block_autogeneration: true

"Found 2026-09-20 at the TOP of the core-complex interactome read queue: 4 baits, 16 IntAct rows, 2 publications, methods including isothermal titration calorimetry, NMR, X-ray crystallography, pull-down and BioID. The evidence is excellent and the inference would be wrong. H3C1 is ONE of many near-identical histone H3 cluster genes, and it appears as an IntAct 'partner' because these proteins bind the H3 TAIL - it is the SUBSTRATE, not a private interactor... If the H3 tail interaction deserves an entry it should be written as the mark or the substrate, not as a histone cluster gene."

This is a live instance of rule 88 - the registry is read too late - in the coordinator's own dispatch. The registry has 147 rows and is working as a place to write things down; the brief that sent this beat out did not consult it, and re-posed a settled question as the beat's centrepiece. The corollary in rule 88 says a beat reporting a trap must say whether the row already existed. It did, with the same date, the same counts and the same verdict.

What is genuinely new here are the two numbers the row leaves open: it says H3C1 is "ONE of many" cluster genes without saying how many, and it does not quantify what counting substrates costs the coverage figure. Both are measured below.

All 16 H3C1 rows in the union resolve to the same UniProt accession, P68431, Histone H3.1, 136 aa, whose gene-names field carries ten gene names 2:

H3C1, H3C2, H3C3, H3C4, H3C6, H3C7, H3C8, H3C10, H3C11, H3C12

The read queue reports H3C1 because the union parser resolved the interactor to a symbol and took the first. Confirmed independently in the raw MITAB: a single PSICQUIC query for KMT2A returns H3C1, H3C2, H3C3, H3C4, H3C6 and H3C7 each appearing 11 times, because they are aliases of one interactor record, not six interactors 2.

So gene-h3c1 would be an entry asserting one of ten equivalent gene records, and a swap among the ten is undetectable by accession, by length, or by any check in this repository. The registry already holds a row for exactly this failure mode at two genes - H3-3A / H3-3B, both mapping to P84243, flagged as unresolvable by accession-plus-length. This is the same class at five times the multiplicity.

Where the 16 rows come from, and which of them is the strong evidence:

Every structural method in the queue's top row belongs to one paper about one module of KMT2A - and that module is deleted from the fusion. PMID 20541251 is about the PHD3-bromodomain cassette, which lies C-terminal to every KMT2A fusion breakpoint and is absent from KMT2A::AFF1. So the strongest evidence in the entire read queue describes a contact the index fusion cannot make.

The counter-argument, stated fairly: the biology is real and important. Proline isomerisation in the PHD3-bromo cassette coupling H3K4me readout to CyP33 and HDAC repression is a genuine, structurally resolved mechanism. It is simply not an interactome membership fact, and this graph records it where it belongs - mark-h3k4me3 and the KMT2A reader-module entries - rather than as a partner row.

Of the 122 union partners with no entry of their own, 11 (9.0%) are histone or ribosomal-protein symbols. Every one of them, with its bait, row count and method 3:

None of the eleven has an entry, so the arithmetic is clean:

The inflation is 2.3 percentage points. That is real and it is small, and it is the honest answer to the question the dispatch posed. Anyone hoping the substrate class was quietly costing the project ten points of apparent coverage should stop: it is not, and the coverage figure is not the reason to exclude them.

The reason to exclude them is the object, and H3C13 is the cleanest demonstration in the dataset. H3C13 enters the union as a partner of DOT1L, by methyltransferase assay. DOT1L is the H3K79 methyltransferase; H3 is its substrate; the assay is the enzyme acting on the substrate. A curator recorded it as an interaction because it is one in the PSI-MI sense, and the union counted it as an unwritten partner. An interactome census that counts enzyme-substrate assays as memberships is measuring something other than membership, and no coverage percentage can repair that.

And this graph's own practice already draws the line correctly. It holds gene-h2ax and gene-h2az1 - two histone variants with distinct, non-redundant biology and their own 1:1 accessions - and no replication-dependent cluster-gene entries at all. Variants get entries; cluster genes do not. That is the right rule and it was in force before this beat.

The registry row's closing sentence - "if the H3 tail interaction deserves an entry it should be written as the mark or the substrate, not as a histone cluster gene" - describes what this graph already does. As of 2026-09-22 it holds nineteen mark-* nodes, of which sixteen are on the H3 tail: H3K4me2, H3K4me3, H3K9ac, H3K9me1, H3K9me2, H3K9me3, H3K27ac, H3K27me3, H3K36me2-3, H3K79me1, H3K79me2, H3K79me3, H3R2me2a, H3R2me2s, plus H2AK119ub, H2BK120ub, H2BK34ub on other tails and H4K16ac, H4R3me2a, H4R3me2s, and - added by this beat - H4K20me2, the first H4K20 state the graph has ever had.

The brief directed this beat to read mark-h3k4me3 and the WDR5 concept entry before deciding. The relevant content is that the H3-tail layer is developed to the level of individual marks with individual writers, readers and erasers, and that the KMT2A reader modules that engage the tail are represented as reader modules. There is nowhere for gene-h3c1 to add anything, and one place it would subtract: it would create an edge labelled "KMT2A binds H3C1" where the truth is "the KMT2A PHD3-bromo cassette, which the fusion does not have, reads H3K4me on the H3 tail."

The rejection therefore stands, restated with its own falsifier. What would reverse it: a report that a specific H3 cluster gene - one of the ten, distinguishable at the nucleotide or the protein level - has a non-redundant role in KMT2A-rearranged leukaemia, such that naming H3C1 rather than H3C4 asserted something. No such report exists, and the ten-gene accession means it would need to be a transcript-level or copy-number claim, not a protein-interaction one.