Expert training data
Demonstrations, reasoning traces and preference data from people who do the work. Built to a specification and reviewed before delivery.
We turn expert work into training signal.
The next gains won’t come from more data. They’ll come from harder, verifiable work.
120+Expert domains
99.2%Acceptance target
72hrsPilot turnaround
1:1Research partnership
Problem
Models plateau on outputs. They improve on reasoning.
Today’s models answer well and work poorly. Real work is multi-step. It has state, tools, constraints and consequences. Most training data captures none of that.
The signal that matters is how experts decide, what they check, and where they stop. It isn’t on the internet. It has to be built.
What we build
One partner for the whole learning loop.
From the first expert demonstration to the millionth agent rollout.
RL environments
Stateful worlds with real tools, hard tasks and verifiable rewards. Resettable, versioned and ready for agents.
Evaluation
Held-out tasks, calibrated rubrics and failure analysis. Know what improved, what didn’t, and why.
Quality is a process, not a checkpoint.See how we work
Environment library
Real work. Repeatable worlds.
Six environments built around the complexity of professional work. Each one has state, tools, a task and a verifier.
Have a workflow in mind?
SWE / 001 · Sandbox
From a failing test to a verified fix.
Fix a duplicate payment caused by retrying a timed-out request. Keep existing behaviour across the repository intact.
TerminalCode editorTest runner
01
async function processPayment(request) {02
const key = request.idempotencyKey;03
const existing = await ledger.find(key);04
05
if (existing) return existing.result;06
07
return ledger.transaction(async (tx) => {08
const result = await charge(request);09
await tx.record(key, result);10
return result;11
});12
}RolloutReward —
Reproduced the duplicate-charge failurePending
Applied idempotency guardPending
24 / 24 regression tests passedPending
Deterministic verification
Custom environments are scoped to your model, tools and evaluation needs.
How we work
More signal. Less guesswork.
Every delivery starts with a definition of success and ends with a measurement against it.
Specify
Start with the gap.Agree on the capability, the task distribution and the acceptance criteria before anything is produced.
Construct
Make the work real.Build tasks, environments and reference solutions with the context an agent actually needs.
Challenge
Test the signal.Review expert work, audit verifiers and hunt for shortcuts that reward the wrong behaviour.
Refine
Close the loop.Inspect model attempts, evaluate on held-out tasks and target the failures that persist.
Expert-reviewed Verifier-tested Version-controlled Evaluation-led
Questions
Built for hard problems.
Different models. Different frontiers. This is how we usually start.
A focused expert dataset, a custom reinforcement learning environment, or a program that spans demonstrations, rewards and evaluation. Start with the capability you want to improve. We shape the task distribution and the delivery around it.
We set acceptance criteria before production, review expert work, test verifiers against shortcuts and keep evaluation separate from training. The goal is measurable improvement on held-out tasks, not a larger pile of labels.
Yes. A custom program starts with permitted uses, access boundaries and delivery requirements. Environments can mirror your tool interfaces and workflows, with resettable state and versioned tasks.
A capability brief, a small reviewed task set, working verification and an evaluation plan. Use the brief builder on this page to send us a short brief. You keep a copy, and we reply by email.
What should your model learn next?
Bring the capability gap. We’ll build the work that closes it.