AI, held to the standard of care.
Simulated patient environments for post-training. Every reward is checked against the clinical guideline, so models learn the standard of care, not shortcuts.
For labs post-training health models. The same environments test any agent before it reaches patients.
Marguerite D., 67Heart failure, type 2 diabetes
Week 1
Week 2
Week 3
Week 4
Week 5
- Agent A · v2.4
- Agent B · v2.3
Agent AStarting follow-up
Agent BStarting follow-up
Guideline checkEvery decision is checked against the guideline that governs it.
Do your agents meet the bar from
- USPSTF
- CDC
- CMS
- ACC/AHA
- ADA
- Joint Commission
- HEDIS
- NCQA
- AHRQ
- IDSA
- WHO
- NICE
Train health models on rewards they can’t game.
Benchmarks test what a model knows. Ordinal tests what it does across a patient’s 30 days, and every reward traces to a written rule.
Held-out exam · patients the model never trained onDecisions that met the guideline, before and after two weeks of training
Before0%
After two weeks0%
The score went upShare of decisions that met the guideline
–Score–
| Category | Score |
|---|---|
| Round 1 | no data |
| Round 2 | no data |
| Round 3 | no data |
| Round 4 | no data |
| Round 5 | no data |
Because the model stopped askingQuestions asked per patient
–Questions–
| Category | Questions |
|---|---|
| Round 1 | no data |
| Round 2 | no data |
| Round 3 | no data |
| Round 4 | no data |
| Round 5 | no data |
Checking each round’s decisions against the rules…
- Plugs into your harness
- A Gymnasium-style API: reset, step, reward. Runs alongside your own models.
- Held-out patients
- Gains are measured on patients the model never trained on: 70% to 85% in two weeks.
- Rewards it can’t game
- Each reward cites its rule and guideline, so benchmaxxing shows up as a flag, not a score.
Every decision, checked against the guideline that governs it.
Patients hold back facts until the agent asks. Each decision gets a verdict and a reward from a written rule that cites its guideline, the same on every run.
Simulated patient · Marguerite D., 67, heart failure · day 11
Guideline check
Escalate rapid weight gainWhen weight rises 2 kg or more in 3 days, escalate to the care team within 1 day.Source: ACC/AHA heart-failure guideline
Waiting for a decision
The same environments test any agent before it reaches patients.
For health systems and agent builders. Compare vendors on your own patients, or check every version before you ship it.
Post-discharge follow-up · each agent run on 500 simulated patients like yours
| Agent | Met the guideline | Harmful mistakes | Yearly savings (est.) | Safe to deploy? |
|---|---|---|---|---|
| Vendor A’s agent | 0% | $0.0–0.0M | Checking… | |
| Vendor B’s agent | 0% | $0.0–0.0M | Checking… | |
| Your team today | 0% | Baseline | — |
Why Vendor B’s agent is not yet safeTwice it logged a rapid weight gain without escalating it. Both patients were admitted with heart failure. Rule: escalate rapid weight gain · ACC/AHA heart-failure guideline