Federated learning gets pitched as a privacy product. It is not. It is a contract about who carries what cost. Done right, the contract is excellent. Done wrong, it is worse than centralisation, because it adds complexity without removing risk.
The contract
AURORA federation says: data stays on the institution's network; compute travels to data; model updates traverse the mesh, signed and audited. The hospital carries the data. The collective carries the model. Nobody carries the wrong thing.
What this gives you
It gives you three things. (1) No central data lake to compromise. (2) A path to training on populations no single institution can see. (3) A regulatory story that's defensible without contortion.
What this costs
It costs you complexity. Federation runtimes are harder to operate than monoliths. DP budgets are easy to mis-configure. Signed updates need a working PKI. We do not pretend otherwise.
The pilot is also where we are most honest about the gap between the theory and the implementation. The runtime is in alpha. The DP budgets are tunable but not yet automatic. The audit log is exhaustive but not yet human-readable. All of these are open issues, in public.