It generates a hypothesis, runs the compute to test it, and tries to falsify it first — then packages the evidence and compounds. We pointed it at canine hemangiosarcoma and human angiosarcoma.
Every hypothesis ships with the cheap, pre-registered test that would kill it — run before anything expensive. The engine falsified its own flagship idea and caught a tumor-purity confound that would have sent us chasing a ghost. Being wrong fast is the feature.
A result isn't a message — it's a portable, verifiable unit: what was run, against what, the directional signal, the confidence, the limitations. Capsules move between machines and people without losing their meaning. Evidence you can audit, carry, and trust on its own.
Every run gives back — open datasets, signature registries, models — and each contribution lowers the cost of the next. A research engine that gets more valuable the more it gives away. Positive-sum by construction, not by goodwill.
With the test that would kill it.
Pluggable GPU lanes — dock, co-fold, MD, omics.
Cheap pre-registered crux, run first.
Portable evidence: signal, confidence, limits.
Publish datasets & models; the next run gets cheaper.