What Onkydra can do today, and what it cannot.

Those are different claims and both are load-bearing. Geneformer-V1-10M serves in-silico gene deletion on Cloud Run (CPU, scale to zero, weights baked in and sha256-verified at build time); a served /knockout returns a real number with its provenance. Licence Apache 2.0, confirmed 2026-08-16 from the model card's own YAML frontmatter rather than from CLAUDE.md, with no gating fields and no commercial restriction. Deploying it validated NOTHING, and it did not rescue CellOracle either. THE RESULT, pre-registered at gfscope_6d5882c8b8f5d236489a9f6a before any forward pass and published as run: against the measured pooled CRISPR TF screen in DIPG17 (GSE267415), the identical column and identical 41 genes CellOracle was scored on, Geneformer scores Spearman -0.023, 95% CI [-0.337, 0.295], where CellOracle scored -0.026, 95% CI [-0.339, 0.293] and FAILED. This arm is recorded NOT TESTABLE rather than failed, which is a demotion and not a reprieve: the pre-registered criterion reads the SIGN of the correlation, and the sign of this readout is retracted by the same artefact that computes it, because 33 of 41 targets come out positive for a reason that has nothing to do with the target. A criterion whose input is disowned returns no verdict, so the pre-registered test was not performed rather than failed. Nothing is withheld by saying so: the signed figure does not clear the bar either, its interval contains zero, and it is a pass under no reading. On MAGNITUDE, which is the quantity that may be quoted once the sign is gone, it is 0.092, 95% CI [-0.231, 0.397], also containing zero; that figure is post hoc and carries no verdict. It is absence of external support, NOT refutation, because a CRISPR screen measures what a cell needs to PROLIFERATE and neither model predicts that. Two nulls also license nothing about in-silico perturbation as a class, and that inference was drawn here until 2026-08-16 and is withdrawn: V1-10M was trained on a corpus that excluded malignant cells, so it is OUT OF DISTRIBUTION on DMG tumour cells and its null may be an out-of-distribution failure rather than an absent signal. Head to head the two do not correspond either: raw Spearman 0.086, 95% CI [-0.237, 0.391], and 0.121, 95% CI [-0.207, 0.425] controlling for detection depth, both intervals containing zero. Reported together because the raw figure alone is not reportable. Top-10 overlap against the measured screen is 3 of 10, where two independent 10-subsets of 41 share 2.44 by chance and 3 or more happens 46% of the time; CellOracle's 4 of 10 happens 18% of the time. Neither is distinguishable from chance, and 2 of Geneformer's 3 (MYBL2, NR4A3) are on the screen's OWN common-essentials list, which is the easiest way into the top 10 of a proliferation screen without predicting anything about this tumour. The reason the two orderings differ is concrete and measured: |effect| against detected fraction runs -0.344 for Geneformer and +0.496 for CellOracle, so their magnitude orderings are detection-driven in OPPOSITE directions. Three further limits a reader needs. 33 of 41 targets return a POSITIVE shift, meaning deletion moves cells TOWARD the anchor centroid, because discarding information drifts an embedding toward the population mean; the sign is not a direction of identity change. The largest magnitudes sit on the least-detected genes, and the three largest are POU2F3 over 42 cells, TAL1 over 30 and GATA2 over 71, of 9,955 OPC-like cells against a floor of 25, with all three in the reported top 10. And there is NO null distribution and NO per-target uncertainty anywhere in the sweep: the per-cell readouts are averaged and discarded, nothing is permuted and nothing is bootstrapped, so the ordering has no calibrated floor and no magnitude in it is known to be distinguishable from no effect. The blocker recorded here until 2026-08-16 is CLEARED: the served reference was subset to 3,000 highly variable genes, exposing only 1,798 with a Geneformer token, which is out-of-distribution input for a rank encoder. It was rebuilt full-gene from GSE210568_RAW.tar to 30,092 genes and 17,745 tokenisable, over the SAME 40,000 cells, verified identical cell-for-cell and label-for-label. The service refuses to compute below 12,000 tokenisable genes rather than computing and caveating. Shared with CellOracle: the same cells, the same cell-state labels, the same dissociation confound and the same detection floor. Not shared: no GRN, no fit to this reference, and a corpus that PROVABLY predates the data, since V1-10M was trained 2021-06 and GSE210568 went public 2022-08-04.