Three capabilities. One loop.

01 — Evals & Benchmarks

Know precisely where your model fails.

Public benchmarks are saturated, contaminated, or both. Real capability measurement needs private, uncontaminated task sets built by people who understand both the domain and the failure patterns of current models.

• Private evaluation suites Expert-authored task sets in coding, professional reasoning, computer use, agentic workflows, and embodied tasks, held out from all training data.
• Rubric & grader design Decomposed scoring criteria that make expert judgment reproducible, with calibration across graders.
• LLM-as-judge pipelines Judge prompts and validation against expert labels, so automated grading is trusted where it should be and audited where it shouldn't.
• Agentic evaluation Evaluator agents that go beyond single-output judging: they re-run failing tasks, perturb inputs and intermediate steps, and follow a failure through the full trajectory to its root cause. A static judge tells you that a run failed; an evaluator agent tells you where and why.
• Failure-mode analysis A taxonomy of why the model fails, not just where: error categories, difficulty bands, and capability gaps mapped to actionable data requirements — built on the root causes our agentic evaluation surfaces.
• Difficulty calibration Pass@k targeting so tasks land in the band that produces training signal, not trivial passes or impossible tasks.

Deliverable format: task sets, rubrics, grading infrastructure, and a written diagnostic report your research team can act on directly.

02 — Model-Breaking Data

Data aimed at a target, not poured into a lake.

Our datasets start from a diagnosis: every example exists because an eval showed the model needs it. That's what makes each one count — where volume-first pipelines produce thousands of examples of what the model already does well, ours target exactly what it doesn't.

• Expert demonstrations Worked tasks from vetted domain experts: full reasoning, tool use, and artifacts, not just final answers.
• SFT & preference data Supervised fine-tuning sets and pairwise/ranked preference data for reward modeling, with rubric-grounded rationales.
• Agentic trajectories Long-horizon, multi-turn trajectories with tool calls, error recovery, and step-level annotations for process supervision.
• Targeted remediation sets Data built directly against a failure-mode map, whether from our evals or your internal ones.

03 — RL Environments

Where agents train against reality.

We build environments that teach long-horizon behavior: realistic tools, realistic state, and reward signals precise enough to learn from — everything static datasets, capped at single-step examples, can't provide.

• Tool-based gyms Sandboxed API and function-calling environments with structured task specs and verifiable end states.
• UI-based gyms Sandboxed replicas of real application interfaces for computer-use and browser-use agents, spanning multi-app workflows.
• Task & reward design Task structures, reward configuration, and step-level and outcome-level programmatic verifiers with assertion-based scoring.
• Difficulty engineering Tasks calibrated to the band where current agents fail frequently, so training produces model-breaking signal rather than saturation.
• Verification harnesses Standardized verifier APIs and QA rubrics so results are reproducible across environment versions.

Deliverable format: portable environment packages with task inventories, verifier suites, and documentation your training infrastructure can consume directly.