Expert-built RL environments beyond software engineering

The work professionals are paid for, rebuilt as environments an agent can learn from. Each ships with a reference solution and a verifier we attacked before you did.

32,000+executable RL environments built by our team
NeurIPS · ICMLwhere our reinforcement learning research was presented

Domains

The agent works in the tools of the trade and delivers the files a professional would.

  • Electronics

    PCB layout, failure cause analysis
  • Architecture

    IFC construction models from drawing sheets
  • Motion design

    Explainer videos, TV ad openings
  • 3D animation

    Character rigging, skinning, animation
  • Games

    Playable Godot scenes with sound
  • Biology

    Binder design, binding kinetics triage

One task, start to finish

Character rig and animation, built with a technical artist.

BlenderIllustration: one game character in five action poses in a 3D viewport
Brief
Rig and skin a game heroine, then animate five moves to match reference videos.
The agent gets
A character mesh and five reference videos.
The agent delivers
A rigged Blender file with five animation clips.
Reward
  • Rigged delivery
  • Run
  • Jump
  • Punch combo
  • Roundhouse kick
  • Front flip
  • Rig holds up
One gate that must pass, then six requirements graded from 0 to 1.
Verifier checked
  • 0.000empty work
  • 1.000reference
  • 9/9broken checks caught
  • 7/7attacks blocked
Models scored
  • Claude Opus 5.50.99
  • GPT-6.1 Sol0.95
  • Muse Wonder 1.00.60
  • Grok 4.70.57
  • GPT-6 Luna0.37

How a task is built

A practising expert is in the first step and the last.

  1. Author

    An expert builds the task from real work and provides the reference solution.

  2. Verify

    We write the verifier in code and check it on known good and known bad work.

  3. Attack

    Agents try to earn reward without doing the work. Every hole they find gets closed.

  4. Expert QA

    The expert reads the model trials and confirms the verdicts match their judgment.

A reward you can train against

A verifier that can be gamed teaches a model to game it. So we try to break ours first, and ship the record of every attempt.

blocked 0passed 0

Fair

Empty work scores 0. Every correct solution passes, not only ours.

Stable

The reference passes on every repeated run.

Precise

Break one check of the reference and only that score drops.

Isolated

Scores come from code run on the delivered files, in a container without network.

Working with us

Start with the sample. Scale only after a pilot holds up on your own evals.

  1. Sample

    We send nine runnable tasks with every probe, attack and trial.

  2. Call

    We walk through the sample and what your model should get better at.

  3. Pilot

    A small batch in the domains you pick, built with practising experts.

  4. Scale

    Volume delivery in the domains that proved useful.

Request samples

All nine tasks, runnable, with reference solutions, verifiers and every trial.

Domains