RL environments
Tasks, scoring and the full training setup, delivered in your format.
- Task suites with checkable answers
- Scoring, rubrics and evaluations
- A ready-to-run package with baseline results
- Physics-simulated environments for robotics
Engineering
RL environments and expert-labeled datasets for AI labs, delivered in the formats you already train in.
What we build
Expert data and labeling
Answer keys and rubrics written by students who study the field.
- Answer keys and reference solutions
- Grading rubrics with documented criteria
- Preference and ranking data
- Second-expert review on every item
Environments and results
Each is a set of tasks with a fixed answer key. Scores are exact-match accuracy on a real test set.
18,254 training rulings · two real test sets
LiveTrade compliance
Classify a product for US import to its 10-digit tariff code. Scored against real customs rulings: partial credit for each leading digit, full credit for the complete code.
Exact 10-digit code on 200 unseen rulings
- Llama 3.3 70B, fine-tuned · published40%
- GPT-5 Thinking · published25%
- DeepSeek V4 Flash, without tools10.7%
On 632 expert-labeled e-commerce listings, the best published agent scores 46.8%. Human experts score 95%.
Answer key: underwriter decisions
In progressCommercial insurance underwriting
Read a commercial submission, decide whether to accept, decline or refer it, and price it. Graded against underwriter decisions.
First results will be published after the initial run.
Answer key: annotated contracts
In progressLegal contract review
Flag every risky clause in a contract and say why. Graded against a copy annotated by legal studies students.
First results will be published after the initial run.
Delivery formats
Training environments
Ready to plug into your existing training setup
Evaluation suites
Scored tasks for measuring a model
Robotics simulation
Physics-simulated tasks for robot learning
Expert datasets
Answer keys, rubrics and reference solutions
Preference data
Comparisons and rankings from domain experts
Your specification
We match whatever format your team already uses
Majors
Each major labels the data only it can judge. Tap a major to see what.
Philosophy
Domain track
Labels
Grades whether an argument is valid, flags fallacies, writes the rubric for ethics cases.
Example task
Rebut this argument. Graded on logic, not tone.
Who we work with
RL environment and evaluation companies
We build the environment and write the domain tasks.
Robotics and physical AI
Simulated tasks, real-world evaluations, data collection.
Frontier labs and data vendors
A vetted team of experts in one field, delivering labeled batches.
Berkeley labs
Pilots with campus research groups.
Project requests
Send the domain, the format and the timeline. We reply with a written plan within a week.