Two cohorts, a hundred infants each, and opposite answers on which sex does worse

The asymmetry in the last row is not cosmetic and is stated in the evidence caveats: one side of this contradiction is a read table and the other is a sentence. The ANZCHOG side is therefore the better-characterised one, which is not the same as the correct one.

Neither cohort is a random sample of infant ALL, and there are at least four ways both results could be honest.

1. Different composition. ANZCHOG's 119 include 38 KMT2A-germline infants, who do much better (5-year EFS 73% vs 33%). If sex correlated with KMT2A status in that cohort, the sex term could be partly carrying fusion status. The JPLSG multivariable analysis is anchored on the KMT2A-r population. These are not the same estimand.
2. Different protocols and eras. Different induction, different transplant policy — JPLSG MLL-10 explicitly restricted HSCT to its high-risk group — and a sex effect mediated by toxicity or by transplant access would differ between them.
3. Different populations. Pharmacogenetic variation differs between the Japanese and the Australian/New Zealand populations, and thiopurine and other drug-handling variants have known ancestry-structured frequencies.
4. Both are subgroup-sized. At ~50-70 patients per arm, a hazard ratio of 2.4 with a lower bound of 1.35 is compatible with a much smaller true effect, and neither analysis appears to have been pre-specified for sex.

The honest reading is that sex is not established as a prognostic factor in infant ALL in either direction, and that any entry in this graph using sex to stratify risk is resting on one of two mutually exclusive findings.

Because of the ANZCHOG hazard ordering. In that cohort's own Cox model the sex term (2.40) is larger than the KMT2A-rearrangement term (2.29), and both clear significance. This graph spends its effort on the fusion, correctly, but a covariate of comparable magnitude that nobody can sign is a standing reminder that fusion status is not the only large term in the model — and that the project's habit of treating clinical variables as background while treating molecular variables as foreground is a choice, not a finding.

It is also a test case for ana-reading-depth. Had both sides been taken from abstracts, this would have entered the graph as 'sex is prognostic in infant ALL' with two supporting citations and no visible problem, because each abstract is individually unremarkable. The contradiction is only visible when the two are put side by side, which is the thing a knowledge graph is supposed to be for and the thing find-contradictions.py cannot do, since it keys on shared PMIDs and these two papers share none.

This entry was written expecting to add two findings. Both were already here.

• concept-anzchog-infant-all-cohort-presenting-features E1 already carried male sex (HR 2.40, 1.35-4.26) ... independent for EFS, with the KMT2A and age hazards beside it.
• concept-pallor-and-anaemia-infant-all-no-infant-figure E4 and concept-risk-group-cut-points-are-administrative-choices-infant-all E3 each already carried female sex to be the significant prognostic factor from JPLSG MLL-10.

Three entries, two opposite claims about the same quantity, written in different cycles by different agents, each correct about its own source, none linked to the others and none aware there was anything to be aware of. No validator flagged it: every item had a resolving citation, a caveat and published counter-evidence, and coverage stood at 100%.

That is the real finding here, and it is about the instrument rather than about infant leukaemia. Per-entry correctness does not aggregate into whole-graph consistency, and this project's quality measures are all per-entry. The partner-provenance audit reached the same conclusion from the opposite direction in the same cycle — every imported result flagged its own import, and no entry showed the pattern of imports — which suggests the gap is structural and not a pair of coincidences.