Request samples

For frontier labs and research-data teams. Access lasts 30 days.

• Rewind: call the next step in a real project. Each sample pauses a project at a decision and grades the model against what the researcher did.
• Research Tree: find the path to a new result. Each sample breaks a recent paper’s main result into problems, down to exam-sized steps.
• SimLab: long research tasks, compressed into RL environments. Agents learn to search a research space efficiently, in a lab simulated from real data and scored against real results.