01 // What we build
Solutions.
Real tasks. Clear evidence. More capable models.
Source, train, evaluate.
- Business dataAuthorized business data, cleaned and validated.Source
- RLVRCode and STEM tasks with executable checks.Train
- Rubric-based RLExpert rubrics for judgment-heavy work.Train
- Agentic RLMulti-step software tasks with outcome graders.Train
- Browser useWeb-app tasks checked against app state.Train
- Robotics RLMuJoCo navigation, contact and control.Train
- SFTExpert reference solutions as training pairs.Train
- Custom evalsHeld-out tasks that measure real capability.Evaluate
Research-grade tasks
From real research to a scored task.
- 01Main problemReal computational work from the expert’s own research
- 02SubproblemsA few connected steps that build toward the answer
- 03Reference solutionA Python solution the expert has run and debugged
- 04TestsChecks that score the result, not the wording
Hard because of the science, never because information is withheld.
- Mathematics
- Statistics
- Physics
- Chemistry
- Biology
- Earth science
- Materials science
- Mechanical
- Robotics
- Software