We intend to publish our results, including the ones that go against us

On a public band-gap dataset split by chemical system across 16 simulated holders, the pooled model reduced test error by 34–40% relative to the best single holder; on a small steels dataset with three holders, by 22–28%. The predefined threshold was 10%. The holders are simulated from one public dataset, not real laboratories.

Scored on each member's own chemistry, the strongest specialist performed worse with the pooled model. A member-specific head on the shared encoder narrows that gap but does not close it.

A constant per-laboratory offset of 0.1 standard deviations reduces the gain but does not eliminate it. Real inter-laboratory disagreement is more complex than a constant offset, and larger offsets are untested.

A copy frozen at the start has 42–57% more error than a model retrained as five further chemistries join. This was measured by adding data rather than by real-world drift, so it is a lower bound.

Yes. One member doubling its update raised the shared model's error by 81% in a four-member pool. The standard defences require the server to see individual updates, which secure aggregation prevents. No defence that operates on masked updates has been built.

Public data contains no failed syntheses. On public successes, pooling helps, and on compounds reported after the training cutoff the model outperforms both a similarity-only and a size-only baseline in both size-matched settings, on all three random splits tested. Testing the central claim requires a partner laboratory's failure records.

Quantum protocols for this kind of hidden (blind) computation exist in theory. Whether they can preserve both pooling privacy and computation privacy at once, for training of this kind, is unresolved, and they require hardware that does not yet exist.