The case for an open, federated infrastructure across neurosurgical disease — and the principles we will not negotiate while building it.
Neurosurgical disease is rare, severe and lifelong. No single hospital sees enough cases to be self-sufficient. AURORA is the answer to that, written out plainly.
MIT at public alpha. No CLA. Every model card, every benchmark, every endpoint published with the code that produced it.
Patient data does not leave the institution. AURORA federates compute to data, never the other way around.
Pinned weights, pinned containers, pinned data hashes. Every result can be re-run from a single command.
Foundation models are powerful and brittle. Mechanistic simulators are robust and blind. AURORA fuses both, deliberately.
Bias and equity checks run inside the training loop, not on a slide deck after release.
Every recommendation explains itself. Every override is logged. AURORA augments judgement; it never replaces it.
AURORA's first eight modules were chosen because they together span the disease landscape: oncology, congenital, biomechanical, fluid-dynamic, vascular-adjacent, and developmental. The next eight are already drafted as RFCs.
The pilot is intentionally narrow. The horizon is not. Once the architecture proves itself across eight diseases of profoundly different character, the cost of adding the ninth, the twentieth and the fortieth drops with every iteration. The point of AURORA is not the modules we ship — it is the modules we make possible.
Some of what is next on the council's table:
We are not building a product. We are seeding an infrastructure — so that twenty years from now, the questions that take a career to answer take a quarter.