Data that shows where frontier models fail at real work.
Expert-built datasets, evaluations and benchmarks for AI labs, from design and engineering tools to first-person work video and real company data.
Four kinds of data models rarely see.
Design and CAD
Original parametric parts, design history and recorded CAD sessions.
View datasetsComputer use
Real tasks in desktop tools, with automatic checkers and agent pass rates.
View datasetsRobotics video
First-person video of skilled manual work, anonymised and described.
View datasetsReal company data
Anonymised codebases and project archives from companies that shipped.
View datasetsFrom specification to delivery.
Specification
You describe the skill, the format and the models you care about.
Samples
We build 5 to 100 items with a datasheet. You review them before any order.
Model results
We run the tasks against current frontier models and share the pass rates.
Delivery
Weekly batches at the agreed volume, by API or repository.
Hard, original, measured.
Frontier-calibrated difficulty
Every task is tested against current models. If they solve it easily, it doesn't ship.
Rights-cleared, original data
Made by our experts or licensed from its owners. Nothing scraped. IP assigned to you.
Fast samples
5 to 100 samples per dataset before any order, so you can judge the data yourself.
Tell us what your models need.
Describe the skill and the format. We'll come back with a sample plan.