Six domains. The same standard in each.

Coding & Software Engineering

Our deepest capability.

Grounded in our peer-reviewed research on AI-assisted software engineering.

• Repo-level tasks Issue resolution, feature implementation, and refactoring in real repositories, not toy snippets.
• Competitive & algorithmic programming Data-structure and algorithm problems with execution- and test-verified solutions, calibrated to the difficulty band where frontier models start failing.
• Code review Expert review data and evals for AI-authored code; our team has published on automated code review at ICSE and on evaluating LLMs as reviewers.
• Long-horizon SWE workflows Multi-step engineering tasks with builds, tests, and debugging loops; process supervision at the step level.
• Terminal & CLI agents Terminal-Bench-style tasks in real shell environments: dependency setup, builds, test runs, and debugging, verified by end-state checks rather than surface pattern matching.
• Test & verifier design PR-aware test generation and assertion-based verification, feeding directly into RL environment design.

Related research: Automated Code Review in Practice (ICSE-SEIP 2025) · Evaluating Large Language Models for Code Review (2025)

STEM & Scientific Reasoning

The proving ground for reasoning.

Mathematics and the sciences are where verifiable, expert-authored data has the highest training value.

• Expert problem sets Mathematics, physics, biology, and chemistry problems authored by graduate- and research-level experts, calibrated to the difficulty band where frontier models stop being reliable.
• Step-level reasoning & proof verification Solutions annotated step by step, with proofs and derivations checked for validity, not just final-answer correctness.
• Verifiable answers Problems designed around checkable ground truth, feeding programmatic verifiers and reward signals for RL.
• Formal math & scientific computing Theorem proving and formal verification (Lean, Coq) where code is the notation for mathematical validity, plus numerical simulation and computational science verified against scientific ground truth.
• Scientific failure-mode analysis Where models silently break: unjustified steps, hallucinated citations, and reasoning that reaches the right answer for the wrong reason.

Professional & Enterprise Reasoning

The work models are furthest from mastering.

Judgment-heavy, context-dependent, poorly represented on the public internet.

• Domain-expert task sets Finance, legal, consulting, accounting, and operations workflows authored and graded by practitioners with real professional experience.
• Document-grounded reasoning Analysis over contracts, filings, reports, and enterprise records: extraction, synthesis, and judgment under ambiguity.
• Workflow-complete tasks Full professional deliverables — a memo, a model, a filing — rather than isolated Q&A, evaluated with rubrics practitioners recognize.

Computer & Browser Use

Where current agents fail most visibly.

And where difficulty calibration matters most.

• CUA trajectories Human demonstrations of realistic computer work across desktop and web applications, with step-level annotation.
• Mock desktop apps & websites Purpose-built replicas of real applications and sites, giving high-fidelity, controllable environments for training and evaluation without production risk or data leakage.
• DOM-oriented web annotation Beyond pixel-level CUA: element-level labeling grounded in the DOM — selectors, accessibility tree, and structured action targets — for reliable, verifiable web interaction data.
• Web + OS combination Tasks that cross the boundary between browser and operating system: downloading a file, opening it in a native app, and acting on the result in one continuous workflow.
• All major operating systems Coverage across macOS, Windows, and Ubuntu, so agents are trained and evaluated on the platforms they actually run on.
• Multi-app workflows Tasks spanning several tools — the coordination layer where single-app benchmarks stop measuring anything useful.
• Training infrastructure & harness A robust harness orchestrating thousands of tasks in parallel across mock environments, with reset, state capture, and programmatic verification built in.

Agentic Reasoning & Tool Use

Long-horizon behavior is a distinct capability.

It needs distinct data.

• Function & tool calling Schema-correct, context-appropriate call sequences with hard negatives where the right answer is not calling a tool.
• Multi-turn planning & error recovery Trajectories where the agent hits failures and recovers — the behavior thinnest in organic data.
• Long-horizon task completion End-to-end tasks with verifiable goal states, structured for both SFT and RL consumption.
• Agent evaluation Trajectory-level evals with step-level root-cause attribution: which action, tool call, or reasoning step sent the run off course.

Related research: Rethinking Code Review in the Age of AI: A Vision for Agentic Code Review (2026)

Robotics & Embodied AI

Physical intelligence runs a few years behind digital.

Which makes its data the scarcest and most valuable.

• Teleoperation data Human-operated demonstrations for manipulation and navigation, with task metadata and quality screening.
• VLA training data Vision-language-action datasets aligning perception, instruction, and motor sequences.
• Physical-task evaluation Structured evals for embodied models: task success criteria, failure taxonomies, and difficulty banding for real-world skills.