How our AI reasons about your health

Every answer follows the same path: from your own words, to a clear set of medical facts, to a careful next step, with a hard rule that anything serious points you to a doctor, never around one.

From your words to a diagnosis

1

You describe it in your own words

You tell us what is wrong in plain words. No forms, no medical terms.

"I've had a throbbing headache and some blurred vision since yesterday, and light is bothering me."
2

We turn it into structured data

We pull the facts out of your message: your main symptom, how it feels, when it started, and what came with it.

complaint: headache
quality: throbbing
onset: ~24 hours ago
associated: blurred vision, photophobia
3

We ask the questions that narrow it down

Each possible cause has signs to check for. We ask you the questions that rule them in or out, starting with the most serious.

"Is this the most severe headache you have ever had?" "Any weakness, numbness, or trouble speaking?" "Have you had a fever or a stiff neck?"
4

We reach a likely answer, and we say when to see a doctor

We reach a most-likely explanation and a next step. When anything needs a clinician, the next step is a clinician, said plainly and not buried.

most likely: migraine with aura
ruled out for now: emergency red flags absent
next step: self-care, with the exact reasons to see a doctor

The safety model

Reasoning is only trustworthy if it fails safely. These rules hold regardless of what the model produces.

The system knows when to hand off

Some things need a licensed clinician: a formal diagnosis, a prescription, an exam. Pymander does not diagnose, prescribe, or replace your doctor. When your message crosses that line, it tells you to see one, and helps you walk in prepared.

Caution is the default

Our safety checks can only make a message more careful, never less. We treat under-warning as the more serious error, so a real concern is never explained away.

Every claim is grounded

The system draws on graded medical evidence and states what it is based on. It does not invent a source, a figure, or a past conversation.

A fixed engine gates every proposal

Deterministic, auditable rules run before anything reaches you, independent of the model. The same input always meets the same guardrails.

Emergencies escalate at once

If your message suggests a medical emergency, the system directs you to emergency services immediately, before anything else.

Your data serves your care only

Your health information is used to provide your care. It is not sold, and it is not used for advertising.

On par with the gold-standard evaluations for medical reasoning

We hold the reasoning beneath every message to the standard used for human physicians.

98.2%
USMLE

On the United States Medical Licensing Examination, the exam a doctor must pass to practice, our clinical AI answered 479 of 488 questions correctly across two independent, fully logged runs.

Read the full evaluation

A benchmark is not a license, and a high score is not a substitute for a clinician. What it establishes is that the reasoning underneath is sound. That is the floor we build on, not the finished work. The measurement that matters most is continuous care held to a high standard over months and years. A benchmark gets you in the door; knowing the line between guidance and medicine, and handing off at it every time, is what keeps people safe.

For questions about our methodology, or a request for the underlying data, write to hello@pymander.app.