calibratedExpert-authored evaluations.
Challenging tasks, grounded reference work, and explicit rubrics. A sharper view of what agents can do—and where they break.
TASKS / RUBRICS / VERIFIERS
HUMAN EXPERTISE. MACHINE PROGRESS.
Worlds for
intelligence.
AI learns by doing. We’re building the environments, evaluations, and expert data that make the experience matter.
An independent lab in formation. An open frontier.
CALIBRATED / RESEARCH INSTRUMENT 001RL.
EXPERIENCEINTELLIGENCE
THE LEARNING LOOPAct. Learn.
Go further.
Go further.
01 / WORLD02 / ACTION03 / FEEDBACK
OUR THESIS
Intelligence needs experience.
Experience needs a world worth learning from.
01 / RESEARCH DIRECTIONS
From expertise
to capability.
Our research agenda connects human knowledge with the systems that teach agents to reason, act, and improve.
Long-horizon trajectories.
Sequences of decisions across files, tools, and applications. Capturing the path from an open-ended goal to a finished work product.
ACTIONS / STATE / RECOVERY
RL environments.
Resettable worlds with meaningful constraints and verifiable outcomes. Places for agents to practice, explore, and learn from feedback.
WORLDS / REWARDS / LEARNING
02 / INSIDE THE LOOP
A world to explore.
A reason to improve.
We’re interested in the whole episode: the starting conditions, the choices along the way, and the evidence of a successful outcome.
ENVIRONMENT ANATOMY / WORLD–001Conceptual architecture
THE SIMULATED WORLD
Context becomes a place to act.
A persistent workspace of documents, applications, and tools. An objective that takes more than one answer. Constraints that make the choices meaningful.
Initial stateVersioned & resettable
Task objectiveExplicit success criteria
Explore the components of a learning episode.
World → action → observation → feedback → learningA research direction, illustrated.
03 / OUR APPROACH
Build the signal.
Study the change.
The question behind our work: what kinds of experience lead to capabilities that transfer beyond the training task?
Start with expert knowledge.
Translate the decisions, constraints, and edge cases practitioners recognize into tasks with substance.
Make success inspectable.
Combine expert-authored rubrics with executable checks. Preserve the evidence behind each assessment.
Look beyond the score.
Study tool use, recovery, and generalization on held-out tasks. Track the cost of the behavior that improves.
THE QUESTIONS AHEAD
Better worlds.
More capable
agents.
We’re building Calibrated around a simple conviction: the design of an agent’s experience matters as much as the difficulty of its task.
- 01 What makes a reward worth optimizing?
- 02 Which experiences teach recovery?
- 03 When does learning transfer?
04 / BUILD WITH US
Make intelligence
your life’s work.
Build worlds agents can learn in. Design evaluations that reveal something new. Turn hard-won expertise into the next training signal.
We’d like to meet researchers, engineers, and domain experts who want to work at this intersection—and partners who share the ambition.
ResearchEnvironment engineeringDomain expertise
Expressions of interest for future collaboration. Specific roles and projects will be shared as they become available.
THE NEXT FRONTIER IS BUILT.
Give intelligence
more to learn from.
Research collaborations. Data partnerships. Shared ambition.