Research
We publish everything, including the results that go against us
Open publication is the strategy, not a concession. The mechanism being public strengthens the custodian position rather than weakening it — participants can verify that we cannot peek — and it means the credibility of this company rests on work anyone can check.
Publications
- SH-QGAN: a split-head quantum generative model for crystal structures
Peer-reviewed. The publication behind the technical thesis — it is domain access and demonstrated capability, not a claimed advantage over classical generative models.
One publication, by one person. That is the honest size of the track record behind this, and it is the constraint the company is most aware of.
Open questions we are working on
These are the live ones, in the order they block the work.
How much does pooling actually improve the model?
In progressSplit a public materials dataset across simulated holders, train each alone, train the pool, measure the gap. The threshold is set in advance: under roughly 10% gain over the best single holder on a realistic chemical-system split, there is no product, and we publish that and stop.
Does the gain survive realistic inter-lab measurement bias?
PlannedReal labs disagree with each other by more than the signal being learned. If pooling collapses under systematic per-lab offsets, that has to surface now rather than at participant three.
How fast does a frozen checkpoint decay against a continuously retrained model?
PlannedIt decides whether access to a live model is worth more than a copy — which is the whole commercial argument, currently an assertion rather than a curve.
Does running local training through the blind delegation loop preserve both guarantees?
OpenGenuinely unresolved, and possibly the research contribution. Stated as open until proven.
Open source
The training and aggregation client will be open source. The precedent is unambiguous: the platform that ran the largest federated pharma consortium was open-sourced and donated to the Linux Foundation by the company that built it, which remained a unicorn. The code is not the asset. The pooled model, the harmonisation across labs and the custodian position are.
There is nothing to release yet. When there is, a reader should be able to reproduce a headline number in under thirty minutes, and the threat model and its limits ship in the repository rather than in a sales deck.
Standing rule
A claim is not an asset until one real run demonstrates it, and a market is not a market until one named buyer has a budget line.
Every number this company publishes will carry a confidence interval, and the privacy overhead gets published even when it is unflattering.