01 / The constraint
Intelligence has run out of internet.
Published data is static and already trained on. Synthetic reasoning only restates what a model already knows. Contract labs return a result with no task context and no training interface. The one input the frontier is short of — scientific data with a verified physical outcome attached — cannot be scraped. It has to be produced.
02 / The loop
Data trains the model. The model designs the next experiment.
A model's failures are the hypotheses worth testing. A rubric turns each one into a protocol with controls, thresholds and a stopping condition. What comes back from the bench — including every failure — returns as reward, as a harder benchmark, and as the next rubric. Two directions, so the returns compound rather than add: every cycle lowers the cost of the next unit of valid data.
03 / The stack
An experiment should return more than a result.
Rubric Engineering
The scoring layer for scientific work: what counts as correct, what counts as causal, and what is worth running at all.
Lab Context Protocol
Instruments that emit their own training data — context, action, deviation, recovery — without binding to one robot body.
Verifier-Driven RL
Post-training where the reward is a physical outcome, so models converge on reality rather than on a loss curve.
04 / Models
Trained on data that was never public.
Cancer World Model
Perturbation, microenvironment and patient response held in one predictive substrate, evaluated against virtual-cell baselines.
MedGenesis
Clinical reasoning under uncertainty — causal inference, evidence synthesis, quantitative judgment — benchmarked against frontier general models.