Clinical agent evaluation / 01

The decision.
And the path.

A correct diagnosis is only part of the story.
Watch what an AI agent asks, examines, orders and treats.

Explore a real run

API-native encountersReplayable action traces

Inside an encounterRecorded, not live

01 / ChooseClinical agent

Grok 4.5

Allowed action

Select an action to see the exchange.

API request → finding

02 / ObserveClinicalGym API

↺ next decision

Every action recorded

03 / EvaluateTrace + case points

ClinicalGym runs the encounter. ClinicalBench compares the runs.

Featured modelGrok 4.5

Emergency-medicine cases—

Encounters closed—

Diagnosis-gated weighted composite—

A real run / 02

One encounter. Many decisions.

Actual action order. No invented transcript.

Loading the saved benchmark…

Recorded encounterGrok 4.5

HistoryExamTestsDiagnosisTreatmentCatalogFinish

—

Select an action

Follow the recorded path

Move through the encounter—

Action types only. Private case text and model reasoning are not exposed.

Published benchmark / 03

The leaderboard, as published.

Model / saved comparisonWeighted composite

What is being compared?

Download the curated snapshot ↓

What the number means / 04

Measure more than the answer.

Outcome

What was diagnosed and treated.

Current: simulation points

Process

What was asked, examined and ordered.

Current: recorded action trace

Safety

What was unsafe—or left undone.

Assessment: actions and omissions

Doctor-authored. Doctor-validated.

Every case was created by a medical doctor and validated by at least two other doctors.

Get in touch / 05

Contact us

Questions, collaborations or benchmark enquiries.

Mariusz Kurman · ClinicalBench