Task samples

Environments

Get in touch

Train for
discovery.

Train for
discovery.

Train for
discovery.

Reinforcement learning environments for healthcare and science.

Reinforcement learning environments for healthcare and science.

Explore the research

Second Opinion · Ankle injury assessment

Where the care breaks down.

Where the care breaks down.

Where the care breaks down.

Nine model configurations interact with the same clinical environment. The results trace each score back to the records, tool calls, and care decisions that produced it.

Nine model configurations interact with the same clinical environment. The results trace each score back to the records, tool calls, and care decisions that produced it.

Task samples

Task samples

Task samples

View all

Second Opinion · Clinical care environment

Assess an ankle injury. Complete the care.

Assess an ankle injury. Complete the care.

Assess an ankle injury. Complete the care.

Months of symptoms. A new injury. An agent must review the evidence, update the chart, and arrange care that the patient can actually use.

Months of symptoms. A new injury. An agent must review the evidence, update the chart, and arrange care that the patient can actually use.

Read the research

Performance and cost

Nine models on the same clinical assessment.

Mean task score / 100
0
20
40
60
GPT-6 AstraFree$0.3$1$3$10$30

Mean recorded API cost · USD; paid endpoints use log scale

52.19/ 100 mean score
$2.11mean cost
The line connects models with the highest observed score at each cost. 8 scored attempts for GPT-6 Astra. Costs exclude grading and infrastructure.

Chimera · Genetic data analysis

Can AI distinguish a genetic association from a misleading pattern?

Can AI distinguish a genetic association from a misleading pattern?

Can AI distinguish a genetic association from a misleading pattern?

Five AI models analyze the same genetic study. They must account for related participants, find an association in the discovery group, and test it independently in the holdout group.

Five AI models analyze the same genetic study. They must account for related participants, find an association in the discovery group, and test it independently in the holdout group.

Read the research →

Vivarium · Task specification

Settle the primary payer. Then bill the visit.

Settle the primary payer. Then bill the visit.

Settle the primary payer. Then bill the visit.

Explore the task →

Environments

Environments

Environments

View all

Get in touch ↗