Get your API key
Contact us with your research or training use case. Credentials and integration details are shared after access is granted.
Request an API key ↘Clinical RL environment as a service
From patient complaints.
To clinical decisions.
An EHR-like simulation your agent can act in - not just answer.
Gather evidence, choose a
next step, and learn from the trajectory.
Environment serviceepisode →
trajectory
01 / Start with a patientComplaint + initial case
context
HistoryFindingsOrders
Only revealed information reaches the agent↓
02 / Act, then observeYour agent ⇄
ClinicalGym
observation → actionfinding ← step(action)Decisions and results become a trajectory↓
03 / Feed the learning loopTrajectory + feedback
→ your trainerUpdated policy → next episode ↺
Granted access · contact us for API credentials and integration details.
01 Patient complaint02 Evidence-seeking
actions03 Clinical decisions04 Feedback for learning
Open showcase dataset
Six recorded encounters. Every step.
GPT-6.1-Sol on acute intermittent porphyria and VIPoma / WDHA, three attempts each. Full transcripts with analyst review of diagnosis and whole-patient care. Recorded runs, not authored illustrations.
Open on Hugging Face ↗Environment / 01
A complaint is the start.
Not the whole case.
An EHR-like encounter unfolds through actions.
The next observation depends on what the agent does.
Medical Q&A
Read the supplied facts.
Produce an answer.
Clinical agent
Decide which facts to obtain.
Choose, observe, revise, act.
The action changes the information available - not just the wording of a reply.
Initial observation / same case for all 8 rollouts
Same start. Different paths. Different returns.
Inspect one path
T1 / Focused assessment
Agent → ClinicalGym
ClinicalGym → agent
Process points-
Revealed decisionsReturn -
Result points
- Diagnosis
- Management
- Budget
- Result total
How this illustration assigns rewards
Process points, per action
Relevant history +2 · vital assessment +3 · focused exam +3 · targeted test +6
No points
Repeated, irrelevant or incorrect actions earn 0 · finish 0
Result points, at closure
Correct diagnosis +10 · appropriate management +10 · budget kept +10
Illustrative point values, not the production reward policy. As in the live environment, each relevant step earns its own points and the final reward depends on the result. Different valid paths can receive the same return.
Group-relative feedback
Eight returns.
One relative signal.
All eight rollouts share an initial case and policy snapshot.
Complete the batch to reveal final returns and group-relative advantages.
Group mean return
Population standard deviation
Aᵢ = (Rᵢ − mean(R)) / (std(R) + ε)
| Trajectory | Process | Result | Return R | Relative A | Group signal |
|---|
Positive advantage means above this batch's mean - not “safe” or “clinically correct.” A GRPO-style trainer can use the relative signal in its clipped policy objective with reference-policy/KL constraints.
Granted-access API / 02
Your policy.
Our clinical environment.
Send actions. Receive observations.
Use authenticated environment endpoints to start an encounter, inspect permitted actions and advance the episode.
Your worker → clinical actionClinicalGym → observation +
feedback
Collect trajectories and returns
Route episode feedback into your training pipeline. Keep held-out cases separate when measuring progress.
Discuss rollout integration →Access is granted, not self-service.
Endpoint URLs, authentication format, quotas and the final reward contract come with onboarding. The batch above illustrates the training flow; it does not call the service.
Policy learning / 03
From one decision.
To a better decision policy.
Start from a complaint
Initial context. Hidden findings. A permitted clinical action space.
o₀
Acquire evidence & act
Questions, examinations, tests, diagnosis and treatment choices.
o[t] → a[t] → o[t+1]
Inspect the whole path
Points per category, missed findings and budget use.
trajectory + scores
Update, then try again
Use the rollout in your RL or preference-learning pipeline.
πθ → πθ′ ↺
What the policy must decide
- What to ask next
- Which examination matters
- When a test adds evidence
- When to revise a diagnosis
- How to treat or stop
A correct answer ≠ a good trajectory
What you reward shapes what the agent does.
State → action → next state
Step-wise observations, actions, rewards and termination.
transitions.jsonl
Compare trajectories
Trajectory pairs with score gaps and behavior diagnostics.
dpo_training.jsonl
Make feedback inspectable
Trajectory records, scoring metadata and filter reports.
reward_model_training.jsonl
You own the training loop.
Online PPO/GRPO training is not bundled. Process and result points, verifier scores and export rewards are research signals, not evidence of clinical safety. Measure progress on held-out encounters with ClinicalBench.
Two layers. Different jobs.
ClinicalGym executes.
ClinicalBench measures.
Build and inspect an agent in the environment.
Compare its performance on ClinicalBench.
Work with us / 04
Bring your agent.
Request API access. Tell us about your model, rollout workload and research use case.
Email us ↗Mariusz Kurman ·
ClinicalGym