Questions about many people, never about one
This door answers population questions from PRAXIS records whose owners said yes to research. It never returns a person, a session or a row — only aggregates that pass the rules below, enforced on the server.
Consent firstOnly records whose owner switched on "Contribute to research" in the PRAXIS app. Switching it off removes them from the next query.
At least 5 peopleAny result, histogram bin or week with fewer than 5 people is suppressed.
Noise on every countLaplace noise (ε = 0.5 per query) on counts; medians are coarsened. Each person counts once, by their own median.
Budget and audit10 queries per researcher per day. Every query is logged with who asked, what, and when.
Ask a question
Synthetic: 64 generated people, open to anyone, for trying the door.
Choose a measure and run a query.
Federated job specification
Models are meant to train where the data lives — on each person's phone and at each site — with only noise-protected model updates leaving. This builder writes the job specification in the form the federated node will accept. Execution is not live yet: it begins in Phase 2 with the federated-learning partner, and every job will need ethics approval before it runs.
Features