
Janus runs OSCE stations with AI patients who hold back clues until asked, and AI examiners who mark against doctors' checklists. Students use it to prepare for their exams. Teams building and buying health AI use it to find the failures before patients do.
Every doctor, nurse and physio passes an Objective Structured Clinical Examination: a patient, a candidate and an examiner with a checklist. Janus recreates each seat.
Built with doctors. Our AI patients hide clues until they're asked, the way real patients do, and react to how they're spoken to.
A student preparing for their exams, or any health chatbot, assistant or agent, plugged in by link. Neither sees the patient's notes.
Marks against doctors' checklists and quotes the transcript for every mark. Miss a danger sign and the station is failed.
Ask an ordinary AI to play a patient and it tells you everything in the first message. That makes every candidate look better than they are.
of replies from popular AI patient simulators were rated “too informative” by clinicians.
Patients who overshare hide the failures real patients expose. Schlager et al., arXiv 2026 (preprint).
Cases, patient behaviour and marking rubrics are written by doctors working in public healthcare, not generated.
The examiner quotes the transcript behind each mark, so you can see exactly how a grade was judged, and dispute it.
We don't build the AI being tested, so we can grade everyone's by the same exam.
Sit a history-taking station as a student, or run your health AI through the same exam.