Make the first product useful to a responsible buyer
The proposed first Discovery engagement is institutional research feasibility: determine whether a resource could support a defined question, through a scoped task and reviewed output. It is a product hypothesis to test with research owners, custodians and budget holders. We have not established willingness to pay or announced a design partner.
Local custody is a starting point
Existing companies already build around distributed health-data research. TriNetX describes a federated real-world data network; Owkin describes institutional federated learning. These examples show an established category. They do not establish a market for Discovery or an exclusive technical advantage.
Discovery’s intended contribution is to connect resource deposits, accountable identities, specific task permissions, bounded proof claims and reviewed releases into one understandable process. Each connection still needs implementation and evidence.
Review the output, as well as the input
The Nature study Disparate privacy risks from medical AI reports that aggregate membership-inference results can conceal substantial risk to individual patients in the models and datasets studied. Our design implication is to evaluate relevant patient-level risks rather than assume that a reassuring average or local training is enough. This is not evidence that Discovery has passed such an evaluation.
A bounded proof proves a bounded statement
A zero-knowledge proof may establish a defined claim without revealing its underlying attributes. It does not establish that a dataset is anonymous, a model is medically correct or a recipient is authorized for every purpose. The exact claim and its release context must remain visible.
A faster answer must still be an accountable answer
A prospective pilot should compare a documented baseline with time to an approved answer, active reviewer effort, reproducibility and negative-case denials. Thresholds must be agreed before evaluation. Discovery publishes no pilot performance claims today.