Reading the same-source counter-evidence: 77 notes adjudicated, and the defect is not the one the metric measures

Is 82.9% independence a real number, or is it propped up by counter-notes that gesture at independence without carrying it?

After 77 careful reads the honest answer is: the number is real, and it is also the wrong number.

Real, because the same-source notes are in good shape. 71 of 77 were legitimate and several were excellent - a paper's own in vivo failure set against its own in vitro headline (target-mcl1-with-bcl2-and-dexamethasone-infant-kmt2a-all E1), a study's own agreement analysis set against the discordant tail the entry was built on (risk-mrd-target-choice-changes-who-is-intensified E1), a diagnostic paper's own three false positives and eight false negatives (concept-flow-cytometry-immunophenotype-kmt2a-r E3). The brief's instinct that same-source opposition is often the strongest opposition available is confirmed by the sample, and nothing in this sweep was deleted.

Wrong, because same-source and independent are not the two categories that matter. The 6 defective notes were all same-source by accident: they stated a fact drawn from some other paper and inherited the claim's citation block because nobody attached a new one. Had the writer attached any second identifier, the same defective note would have counted toward the 82.9%. The metric rewards the motion that produces the error.

So the recommendation is not 'reduce same-source notes'. It is: measure whether a counter-note's citation supports its own text, which nothing in this repo currently does.

Two things about the boundary between 1 and 2, because the brief's definition of category 1 is narrower than what the graph actually contains.

Most category-1 notes are not the authors' stated limitation. They are the reviewing agent's reading of the paper's own data against the entry's interpretation. That is a stronger form of same-source opposition than a quoted limitations paragraph, not a weaker one, but it is a different thing and the notes rarely say which they are. Where a note IS the authors' own statement - cell-state-pro-b-arrest E_jansen, risk-mrd-target-choice-changes-who-is-intensified E2 - it is worth one word to say so, because a reader weighs 'the authors concede' differently from 'the reviewer infers'.

The category-2 population is dominated by one shape, repeated across seven fusion entries: the KMT2A recombinome (PMID 37019990) is a referral series, so every percentage taken from it is a share of referrals rather than of the disease. That argument cannot be made with a second paper, because no population-based partner-distribution comparator is published; it is legitimate by necessity and it is correctly stated. The sweep checked Interfant-06 (PMID 31283407) as a prospective comparator and the published abstract does not give a partner breakdown, so nothing was added.

Each repair adds a note alongside the original; nothing was removed, and every identifier below was fetched from PubMed or the primary source in this session rather than copied from a sibling entry.

1. drug-silmitasertib E2 - the worst of the six, and it was in a caveat rather than a note. The entry argued for hitting CKII instead of taspase1 because a 30-million-compound virtual screen against taspase1 failed, and its caveat said this was 'the only substantive public information on taspase1's druggability'. That is false. PMID 22166309 (Chen et al., Cancer Res 2012) reports NSC48300, a reversible non-competitive taspase1 inhibitor at Ki = 4.22 uM that inhibited tumour growth in MMTV-neu and U251 models, and notes that Taspase1-null mice have no overt deficit. PMID 38958349 (Hommel et al., Chem Eur J 2024) reports multivalent molecular tweezers against taspase1's surface loop at nanomolar affinity with efficient proteolytic inhibition. Neither was cited anywhere in this graph. The caveat was corrected and both papers added as counter-evidence with their limits stated (micromolar arsonic acid, no leukaemia model; supramolecular chemistry, no cellular cleavage assay).

2. mechanism-kmt2a-subnuclear-localization E5 - the graph cited the right paper in ten other entries. The note argued, correctly, that 'SEC associates with every chimera tested' does not generalise across partner classes, and supported it with 'a 15-residue gephyrin tubulin-binding domain from a protein with no transcriptional activity is sufficient to transform' - under PMID 20159561, the SEC purification paper, which does not contain that result. It is PMID 14751928 (Eguchi, Eguchi-Ishimae & Greaves, Blood 2004), The small oligomerization domain of gephyrin converts MLL to an oncogene, and this graph already cites it in ten entries. This is the mirror of the gene-ssrp1 failure Rule 64 was written for: there, an absence was declared while the disproving paper was cited four times elsewhere; here a finding was asserted while its source was cited ten times elsewhere.

3. cell-type-fetal-liver-hsc E1 - and the independent literature turns out to disagree with itself. The note asserted 'LIN28B is required for KMT2A::AFF1 to transform human fetal liver CD34+ cells at all' under PMID 30679323, which contains no LIN28B experiment. The source is PMID 42213620 (Ling et al., Blood), cited once elsewhere in the graph but not here. Fetching it surfaced something the same-source note could never have shown: PMID 36763502 (Eldeeb et al., Cell Rep 2023) reports that in an MLL-ENL/LIN28B dual-transgenic mouse LIN28B impedes MLL-rearranged AML through suppressed MYB signalling and proposes fetal LIN28B as protection against MYB-dependent AML. That paper was cited nowhere in this graph. Required in human fetal B-lymphoid transformation, protective in mouse myeloid disease - both were added, with the lineage/species caveat that the disagreement may not be a contradiction.

4. risk-menin-pocket-occupancy-endocrine-late-effects E1 - a counter-note that went stale when its own entry was corrected. The note ended 'Nothing published shows it does.' The entry itself carries a CORRECTION section and an E3 recording that the REVUFORJ label reports intact PTH increased in 34% of 241 patients and lists hyperparathyroidism under Endocrine disorders. The correction reached the summary, the body and the evidence array, and never reached counter_evidence. The label was re-fetched from DailyMed on 2026-09-20 (34%, N=241, denominators 139-240; phosphate increased 51% / decreased 25%; corrected calcium increased 15%, 0% grade 3-4; hyperparathyroidism under Endocrine disorders, <20%) and added as a second note, deliberately not by copying the sibling entry's identifier block.

5. target-osteoclast-inhibition-zoledronic-acid-high-risk-b-all E1 - the same-source note was true and understated the confound. Full text of PMC11868389 confirms the paper itself states dasatinib inhibits osteoclast formation and activity, and reports zoledronic acid as more potent - but attributes the dasatinib effect to two prior papers that were not cited. The primary one, PMID 20225261 (Vandyke et al., JBMR 2010), shows dasatinib alone raising trabecular bone volume and thickness in rats at levels comparable with the zoledronic-acid arm, with reduced osteoclast number, surface and serum CTX-1. Cited independently, the confound in the BCR::ABL1 combination arms is larger than the index paper's supplementary comparison implies.

6. cell-state-under-menin-inhibition E3 - an uncited 'well documented'. The note argued that the revumenib lineage-switch observation may be a general dose-of-oncoprotein effect, on the basis that 'lineage switch in KMT2A-rearranged leukaemia is well documented after CD19-directed therapy' - with no citation. PMID 26914337 (Rayes, McMasters & O'Brien, Pediatr Blood Cancer 2016) reports an infant with t(4;11) ALL who switched to CD19-negative monoblastic AML on blinatumomab: index fusion, index age, and a drug that does not touch menin. It was cited nowhere in the graph. The bounding query is recorded in the new note.

Going in, the expected failure mode was the one the brief describes - a lazy same-source note where a second study existed and nobody looked. That is not what the sample contains. In five of the six gaps the writer HAD read independent literature; the fact in the note is correct and the reasoning is good. What is missing is the identifier. The note inherited the claim's citation block because the counter-evidence array has no convention for 'this note's source differs from its parent's'.

This produces a defect with a specific and nasty signature:

• verify-citations.py passes it by construction. The stored PMID and the stored title agree with each other and with PubMed. The script's whole design is to catch a title that describes a different paper; here the title describes exactly the right paper, cited for something it does not say.
• verify-claims.py cannot see it. It scores each evidence item's claim against its abstract. counter_evidence[].note is not in its scope at all.
• find-contradictions.py cannot see it. The note is not contradicting a sibling entry; it is quietly unsupported.
• And the independence metric scores it as a virtue-adjacent conservatism. A same-source note reads as cautious. This one is a citation error wearing caution's clothes.

The cheap detector, if a future cycle wants one: for every counter_evidence[] whose PMID equals its parent evidence item's PMID, score the note text against that PMID's abstract the way verify-claims.py scores a claim, and rank the worst. All six gaps here would rank badly, because each note's decisive sentence is about a molecule or a result the abstract never mentions. That is the same instrument this project already has, pointed at a field it was never pointed at.

A second thing I did not expect: 39 truncated titles. A mechanical scan of every same-PMID counter-note found 39 whose stored title is a strict PREFIX of the correct title - the listing-truncation trap the brief documents, sitting inside counter-evidence citation blocks where no checker looks. verify-citations.py passes them because an idf-weighted overlap on the first 80-130 characters is high. Two were fixed in cell-type-fetal-liver-hsc during this sweep, from a title fetched in-session; the remaining 39 are listed below for the coordinator, because fixing them from the sibling evidence item's title would be exactly the second-hand-identifier motion the brief forbids.

Recorded so the decisions are re-runnable, per the standing bounded-rejection rule.

Every row is a counter_evidence[].citation.title that is a strict prefix of the same block's correct title (taken as the parent evidence item's title, which verify-citations.py validates against PubMed). Each needs the title re-fetched, not copied.

breakpoint-kmt2a-spliced-fusion-partner-exon-2 E3 (37019990) - concept-bruising-petechiae-thrombocytopenia E2 (34054042) - concept-endocrine-endpoint-absence-menin-trial-registries E4 (39121437) - concept-flow-cytometry-immunophenotype-kmt2a-r E3 (12886247, 82 of 180 chars) - concept-men1-loss-to-endocrine-tumour-latency E3 (17590169), E6 (25594862) - concept-phase2-go-no-go-register E7 (9474863) - cost-frontline-infant-leukemia-care E2 (32589365) - disease-infant-all E1 (31283407) - endocrine-men1-pancreatic-neuroendocrine-tumour E1 (17590169), E2 (25594862) - endocrine-men1-pituitary-adenoma E2 (25594862) - endocrine-men1-primary-hyperparathyroidism E2 (25594862), E4 (9683585) - epi-men1-age-penetrance-curve E2 (9683585) - gene-rnf20 E3 (28690313), E5 (41533567) - gene-ube2a E5 (32415735) - gene-wac E3 (41533567), E5 (26757981) - hyp-h2bub-direction-is-unknown-in-kmt2a E2 (28690313) - hyp-menin-inhibition-perturbs-parathyroid-axis-dose-dependently E3 (17590169) - mechanism-condensates-load-bearing-or-correlative E7 (35073946) - mechanism-fusion-dna-targeting-cxxc-athooks E4 (28076791) - mechanism-hoxa9-meis1-output-program E5 (32242051), E7 (34088716) - platform-trk-inhibitor-tumour-agnostic-precedent E2 (29606586), E3 (31838007) - pt-story-infant-all-kmt2a-aff1-hyperleukocytosis E1 (39717192, reversed - the EVIDENCE title is the short one here) - reg-breakthrough-and-accelerated-approval E11 (40136298) - resistance-cross-resistance-between-menin-inhibitors E3 (41351513) - resistance-enl-marked-epigenetic-memory E3 (40536420) - resistance-men1-interface-mutations E2 (41351513) - resistance-tp53-apoptotic-priming-menin E3 (41351513) - risk-menin-inhibitor-exposure-duration-vs-men1-latency E5 (36552783, 54 of 221 chars) - risk-revumenib-intact-pth-elevation E2 (39121437), E3 (17590169) - subtraction-kmt2a-tad-cbp-contact E4 (30573454) - subtraction-kmt2a-taspase1-cleavage-sites E8 (30573454).

Also found while scanning: 3 counter_evidence items are bare strings rather than objects, and 15 are objects with no citation at all. Those 18 are invisible to every identifier check in the repo and are not counted anywhere in the 6,412/5,317/1,092 census.

Four concrete things, in order of how much they would have caught here.

1. A counter-note that cites its parent's source must say why. One of three words at the front: the authors' own caveat, the paper's own data, or scope. The sweep added that framing where it repaired a note. It costs a phrase and it tells a reader whether the opposition is conceded by the paper or inferred by us.
2. A counter-note that states a fact from elsewhere must carry its own identifier. All six gaps are this rule being broken. The array already supports a per-note citation; nothing enforces that the note's content matches it.
3. Score counter_evidence[].note against its own PMID's abstract, as a small extension of verify-claims.py. This is the only mechanical check that would have found any of the six.
4. When an entry is corrected, sweep its counter_evidence array. risk-menin-pocket-occupancy-endocrine-late-effects is the worked example: an entry that carries a prominent CORRECTION section refuting 'nobody has looked', with a counter-note three fields away still ending 'Nothing published shows it does.' The existing rule do not leave an entry arguing with itself was written for the body contradicting the counter-evidence; this is the same defect in the other direction and the rule should say so.