We are building
medical superintelligence.

Almost everything medicine knows is written down somewhere, and almost none of it is written down in a form a model can learn from. The trial literature is public and thin. The practice, the billions of encounters a year in which clinicians decide under uncertainty and then find out what happened, is private and mostly unrecorded. That is why medical AI has stalled at the level of an impressively well-read student: it can recite the guideline, but nobody has shown it a few million instances of deciding, with the answer attached.

The second problem is reward. Reinforcement learning carried coding and mathematics a long way because both have cheap, honest verifiers, and you cannot unit-test a differential diagnosis. So we manufacture the verifier: real patient trajectories turned into gradeable environments, with specialist-written rubrics where judgement is required and the recorded outcome as the anchor wherever the world settled who was right.

Both problems need a position inside the institutions where care happens, which is why we build the record system too. It will not arrive as a single model release. It arrives as a loop that keeps closing faster: care producing data, data producing models, models returning to care.