Patient simulator
Plays real patients, with messy wording.
Checked against clinicians
Realism against held-out real transcripts; clinician Turing test; check simulated-vs-real performance gap.
Where a person decides
Never used alone for sign-off (Oxford gap).
Works on
Chat, triage, voice, mental health.
In development. Early access is open. Reviewed .
The problem it solves
Real patients write messy messages. Testing with neat scripted prompts overstates how well a chatbot will do.
How it works
Take a clinical case
A case defines the true condition and the facts a patient would know.
Play the patient
The agent writes it many ways: vague, anxious, with wrong terms, in other languages.
Score the conversation
Outputs are graded for triage correctness and safe advice.
An example, illustrative
'I have chest pain radiating to my left arm.'
'my chest feels tight and my arm is weird, probably nothing?'
Clinician-written cases and the chat endpoint to test.
Simulated patients differ from real ones. We measure that gap rather than assume it away.
What it helps measure
Quick answers
What does the patient simulator do?
Generate realistic, wording-varied patient and caregiver conversations.
How is the patient simulator checked against clinicians?
Realism against held-out real transcripts; clinician Turing test; check simulated-vs-real performance gap.
Does simulation replace real users?
No. It tests many more phrasings cheaply, then real-user checks confirm.
How realistic is it?
We compare simulated and real conversations and report the gap.
What can the patient simulator not do?
Never used alone for sign-off (Oxford gap). No agent makes a clinical decision.