This graph states 624 falsifiers and records the outcome of 17 - a figure this entry first published as 8, because the detector did not know the corpus's own ALL-CAPS convention for recording one

Cycle 60's headline cost one beat. The entry `mechanism-kdm3b-depletion-lowers-ctcf-only-at-the-kmt2a-anchor` carried a field naming the anchor-matched comparison that would refute it, stating the refutation condition numerically, predicting from the paper's own 17,077 losing peaks that refutation was the **likely** outcome, and ending *"NOT RUN by this beat"*. It sat two cycles while the claim stayed published. Run, it killed the claim and confirmed the rival. Nothing in this project reads those fields, so this sweep was named in the cycle-60 log as the highest expected value per unit cost here. **THE FIELD THAT PROMPTED IT IS A ONE-OFF.** `the_check_that_would_refute_it` appears in **one** entry. The corpus convention is a `falsifiers` list, on **161** entries, plus about two dozen assorted `kill_criterion` / `decisive_*` / `falsifiable_*` fields. A one-off field name is exactly what no sweep catches, which is why that finding sat unbought. **MEASURED over all 2,125 entries and 10,771 edges from 280 edge files: 161 entries state a falsifier and 1,964 state none. Of 624 stated falsifier items, 610 record no outcome, 8 record one, and 6 say explicitly that the test was not run.** The queue of unrun-or-unrecorded falsifiers on entries not already hedged to weak or speculative is **118**, and there are **21 entries where every stated falsifier is unrecorded** and at least two exist. **THE CLAIM THIS LICENSES IS NARROWER THAN 8-OF-624, AND I PROVED IT BY HAND.** A keyword sweep sees whether a *field records* an outcome, not whether a *test happened*. Ranked by edge exposure the top of the queue is `mechanism-h3k9me1-2-ctcf-balance-gates-kmt2a-rearrangement` - exposure **38**, confidence moderate, status verified, **six** falsifiers, none recording an outcome, and the entry carrying this graph's prophylactic-intervention proposal. I read all six. **One had already been run, in cycle 59, and the entry never said so.** So the real run rate is higher than 8/624 and unknown. What is measured is the **recording** rate, and that is the defect: this corpus states how it could be wrong and does not come back to say what happened. **AND THE ONE THAT WAS RUN SPLITS BY SCALE, WHICH IS A DEFECT IN THE FALSIFIER.** It asked whether H3K9me3 moves with me1/me2 at the BCR in GSE210480. Re-read from `analyses/q108-gse210480-h3k9me3-kmt2a-bcr/results.json` rather than from entry prose: at the 1,048 bp CTCF element H3K9me3 is **1.014** while H3K9me1 is **3.738**, and at exon 11 and the summit 1.042 against 3.852 - so it **fails to fire** there. Over the 7,466 bp BCR core me3 rises **1.344** against me1's 2.675, ranking **3,877 of 18,082** chr11 bins against me1's **72** - so it **partly fires** there. The falsifier names no interval, so it can be met or not met depending on a scale it never specified. **The mark that tracks me1 least is me2, which is in that entry's own title.** Over the BCR core H3K9me2's log2 is **-0.0326**, rank **10,276 of 18,082** - the middle of chromosome 11 - and at the element a sibling entry records me2 **falling to 0.44x** while me1 rises 3.9x. The entry is named for an "H3K9me1/2 balance"; the measurement is me1 up, me2 down, me3 slightly up. **No falsifier in that entry tests its own titular pairing.** Score on that entry: **1 of 6 run**, two of the remaining five being bench experiments this project could specify today - and one of those is already written up as `exp-h3k9me3-and-ctcf-at-the-kmt2a-anchor-in-a-cd34-progenitor`. Its confidence was **not** lowered, because nothing in the adjudication refutes the mechanism. What was removed is its implicit claim to having been tested. **CORRECTED IN CYCLE 62, AND THE CORRECTION IS THE MOST INSTRUCTIVE PART.** This entry first reported **8** recorded outcomes. Hand-reading a second queue entry - `hyp-obligate-partner-interface-generalises`, 21 edges, eight falsifiers - showed two of them begin ***"ALREADY OBSERVED"***, and `hyp-nup98-partner-obligate-interface` carries ***"ALREADY PARTLY RESOLVED IN THE WRONG DIRECTION"*** and *"is a category error as written"*. **This corpus does have a convention for recording a cashed falsifier — an ALL-CAPS status phrase leading the text — and the detector did not know it.** Corrected: **17 RUN, 601 UNMARKED, 6 NOT_RUN**, and the queue falls from 118 to **113**. **The direction survives and the number did not.** 17 of 624 is 2.7%, so *almost never* still holds — but the published figure was wrong by more than a factor of two and it was in this entry's title, in a commit title and in the ledger. The id still reads `...-outcome-of-eight` and is deliberately not renamed, because ids carry edges. **And the instrument is now wrong three times out of three on every case anyone checked.** Three entries have been hand-read across cycles 61 and 62 and every one recorded outcomes scored as unrecorded. That is the real result: the UNMARKED class means *no outcome detected*, not *no outcome recorded*, and the 601 is an upper bound whose true value nobody knows. What is solid is the shape — a corpus that states falsifiers far more readily than it returns to them — and the 1,977 entries that state none at all, which no detector error can move.