Why a result in a model is not a result in a patient.

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also license nothing about in-silico perturbation as a class, and that inference was drawn here until 2026-08-16 and is withdrawn: V1-10M was trained on a corpus that excluded malignant cells, so it is OUT OF DISTRIBUTION on DMG tumour cells and its null may be an out-of-distribution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.

Floor: zero is the modal outcome, so a zero cannot be read as a negative finding. Clears when: A network with more of these targets as regulators, or an honest statement that the target is out of scope. The second is what the product does now.

Clears when: More lines. At six, the exact null is so wide that only a perfect ordering is nominally significant.

Clears when: Depth-matched libraries, more animals per arm, or a composition readout that is not a mixture mean.

Floor: the measured repeatability of the one readout that measures its own. Clears when: A combination arm measured alongside both single agents, with intervals, and a named reference model. Bliss needs a fraction affected and Loewe needs a concentration, and no held readout has either.

Within the stronger one, the three patient tumours disagree with each other. 3 of 6 models carry identical curated mutation sets and cannot be separated by the ranking at all. Floor: exact, over all 720 orderings of six models. Clears when: An outcome attached to a model choice. Nothing in the prediction ledger has come back, so there is no such pair anywhere yet.

The genetic-dependency readout separates cleanly by comparison, at 0.6662 against 0.8616 over 352 genes. In decision units the same inversion holds: the drug cross-model flip rate is 0.0768 against a same-model noise floor of 0.1051. Floor: 10,000 random orderings, cleared by both readouts and not the binding constraint here. Clears when: A drug screen with replicate plates rather than repeated library entries. The repeat figure is a LOWER bound on this screen's repeatability, because the two entries come through different compound libraries, so a fair repeatability could put the ceiling back above the cross-model figure and restore real heterogeneity. Undemonstrable, not absent.

Today the numbers are 0.635 against 0.4349, so the condition would not fire; the point is that nothing would happen if they were reversed.

Which programme is the headline is a biological judgement about the indication, not a default. There is no rule that picks one from an arbitrary reference, and picking the largest cluster would pick `Other`, which is the residue. What would remove it: A second indication's single-cell reference with its own annotated malignant compartments, plus a stated reason for which of them is the readout. The name would then move onto the atlas release rather than into a module constant.

That is a stronger provenance than a clustering this product invented, and it is exactly why it does not transfer: the vocabulary belongs to that deposit's authors, and another indication's deposit will have named different things. What would remove it: An author-annotated reference for the second indication. Absent one, the alternative is clustering its cells here and naming the clusters, which is a scientific claim this product is not in a position to make and would have to be labelled as one.

That divergence is a correctness risk today and is deliberately NOT fixed here, because every one of those files produced a checked-in artefact and editing a marker list without regenerating would change what a served number means while leaving the number in place. What would remove it: Nothing about a second indication removes this one. It needs the six copies collapsed into a single artefact-carried vocabulary and every dependent artefact regenerated on the same day, which is a scoped piece of work on the DMG side and a prerequisite for the second indication rather than part of it.

It has to be exported, because `provenance.markers` reported the INTENDED lists until 2026-08-06 and every surface consequently overstated what the numbers were made of. The AC-like programme is two genes. What would remove it: The generated table gaining an indication key, which is downstream of the atlas release gaining one, which is downstream of a second atlas existing.

A bridge edge is an explicit modelling assumption and the evidence licensing it is disease-specific by construction. Five further entries abstain on DMG-specific literature, so the abstentions are as coupled as the enablement. What would remove it: Equivalent in vivo pharmacodynamic evidence in the second indication, per edge. Not transferable: an edge is enabled by a study, and there is no study about two diseases at once.

The axis is real and it is worth having: ACVR1 reads 0.805 in H3.1 against 0.035 in H3.3, so ignoring it would pool two different diseases. What would remove it: The second indication's own molecular axis, whatever it is, plus a transfer-relevant measurement on it. The shape of the mechanism generalises; the axis does not, and inventing one would be a claim rather than a parameterisation.

The exclusion is correct and it is specific: the circular genes are whichever genes define that indication's programmes. What would remove it: The second indication's marker sets, from which its own exclusions are derived. The exclusion RULE is general and is already written as one; only the gene list is coupled.

The three are chosen because of what they showed on this artefact, and OLIG1's row is the one that retracted a claim (leave-target-out exactly 0.0000 in all four cell states, so its whole effect was its own expression averaged into a programme it is a marker of). What would remove it: The second indication's own lineage regulators and its own handling control. Note the control half is harder than it looks: FOS and JUNB are unusable as comparators here because they are handling-associated on this reference, measured at Cliff's delta 0.63 between single-cell and single-nuclei libraries.

These are the filenames of files that exist, pinned by sha256 in the checksum registry, and renaming them to a pattern would break every pin for no benefit while there is one indication. What would remove it: A second indication's equivalent artefacts existing on disk, at which point the graph takes an indication and composes the name. Doing it before they exist would produce a lookup that resolves to nothing and a code path nobody can run.

The gate mechanism is general; its contents are a record of what has been served, and a record cannot be about a thing that has not happened. What would remove it: Nothing. This entry exists so the count is honest, not because it is fixable: the register fills as methods land, and a second indication's methods will add rows rather than change these.

These are the specific things a writer would otherwise overclaim about this evidence, so a generic version would be a weaker critic. What would remove it: Per-indication critic instructions, which need a second indication's evidence to be written against. Generalising the wording without the evidence would delete a working check and replace it with a placeholder.

The caveat corpus itself is DMG-only, so a second indication gets no caveats rather than the wrong ones. What would remove it: Its own caveat corpus per indication. The mechanism is already right and this is the shape the rest should copy.

The answer, 30 midline against 11 hemispheric with 6 unstated and 5 ependymoma comparators, is why the module exists. What would remove it: The second indication's own site vocabulary and its own composition question. Anatomical site is not even the relevant axis for most cancers, so this is a coupling to the disease's anatomy rather than to its data.

The only defence is application-level: the coordinator overwrites whatever the planner produced with the library's own slug. Recorded rather than fixed, because a check constraint enumerating today's keys is a constraint that has to be migrated every time an indication is added, and it would encode the registry in a place the registry cannot read. What would remove it: An indications table the eight columns reference. Deliberately NOT written as a migration today, and the reason is the same reason it is recorded here rather than fixed: a reference table with one row constrains nothing, and a check constraint enumerating today's keys would have to be migrated every time an indication is added while encoding the registry somewhere the registry cannot read. It becomes worth doing when there is a second row to put in it, which is the same moment the constraint starts catching something.

Neither carries a floor. The script contains no permutation, no shuffle, no random draw and no baseline of any kind, so r = 0.587 has no zero: both vectors are dominated by the same few high-frequency drivers, and a correlation between two such vectors is high before any biology enters. The earlier r = 0.97 was retracted for circularity, which is the same class of error one level up. What would clear it: Add a permuted-gene-assignment arm to validate_external_cohort.py and report its 95th percentile beside each r, plus a second comparator that is not the anchor at all, such as a generic paediatric high-grade glioma frequency vector. If r = 0.587 does not clear the permuted null on n=18, that is the answer and it goes in the failures register.

What would clear it: Implement the remediation already written into the grandfathered entry: stateCompositionSeries refuses to render a composition when that cohort's transfer agreement is below its own majority-class floor, rather than rendering it beside a note. Today the numbers are 0.635 against 0.4349, so the condition would not fire.

The comparator half does not resolve. The cheapest alternative is one bit per gene, is-it-a-core-housekeeping-gene, with no data, no model and no screen, and it reaches a median AUROC of 0.7431 for the held-out line's strongest dependency decile. Spearman over 352 genes and AUROC over a top decile are different quantities, so 0.7806 against 0.7431 is not a margin. The replicate ceiling is 0.8616, which leaves little room for one to exist. What would clear it: Score the housekeeping flag and the five-line mean on the same metric, both ways: Spearman for the flag over all 352 genes, and AUROC for the mean over the same top decile. Two numbers on one scale, reported whichever way they fall. If the one-bit flag matches the five-line mean, the honest conclusion is that the leave-one-out arm is measuring pan-essentiality, which its own leakage note already raises.

It did not clear its bar. The result is the absence of external support rather than evidence against, because it compares a predicted cell-state shift against a measured proliferation dependency and 22 of its 41 points are structural zeros. It is not claimed as validation anywhere and must not be. Recording it as met on process while it failed on result is exactly the upgrade the honesty invariants forbid. What would clear it: An external check whose readout is the same quantity the model predicts. A CRISPR screen measures what a cell needs to proliferate; the knockout predicts a cell-state programme shift. Until a perturbation dataset with a state readout is ingested, this gate cannot be cleared by trying harder against the screen already held.

ODbL adds share-alike and a positive §4.6 obligation to offer recipients either the derivative database or the method that made it. Whether backend/cohorts/dmg-matrix.ts is a Derivative Database under §4.4b or only a Produced Work under §4.5b is open, and §4.6 catches a Produced Work derived from a Derivative Database too, so the second route only helps if the first does not apply. The licence table records the obligation as attribution yes, share-alike unresolved. Exposure starts at the first paying customer, so this is currently latent rather than absent. What would clear it: A determination on whether the n=60 anchor matrix is a Derivative Database, from someone qualified to make it, recorded in DATASET-LICENCES.md with its reasoning. If it is, publishing the derivation method alongside the product satisfies §4.6 and the gate clears; the method is already public, so the cost is procedural rather than strategic.

This is the deepest of the fourteen and it would hold expansion on its own, because a second indication built the same way inherits the same ceiling. What would clear it: An outcome layer for the indication, from a source that is rights-cleared for commercial use. This is the item on the list most likely to be impossible from open data alone for a rare paediatric cancer, which is a reason to establish it before a second indication is chosen rather than after.

NCI PDMR returns no data for DIPG, midline, pontine, brain or K27. ITCC-P4's 19 are the largest set and are behind an access gate. Direct engraftment of DMG autopsy tissue has also produced histologically convincing pontine tumours composed of MURINE cells, so a record without a species-of-origin check cannot support a drug claim even where one exists. What would clear it: Either an access agreement for a characterised in vivo panel, or a documented decision that the product's recommendations stop at in vitro for this indication and say so on the surface that makes them. The second is cheaper and is a real option, but it has to be stated rather than left as the shape of what exists.

Three surfaces asserted a customer base in the present tense anyway and were retracted on that date. With a zero denominator no per-customer quantity is computed anywhere, which is why the productisation economics artefact reports null rather than zero for every rate. What would clear it: One external account that asks for a named second indication. Note the ordering this gate encodes: the buyer comes before the build, not after it. Building an indication and then looking for someone who wants it is the expensive order and it is the default one.