RL environments · evals · post-training

Train AI on what your experts know, not what it already gets right.

Experts watch frontier models do real work on a real computer and correct them as they go, or do the work themselves while it's recorded. Every correction becomes deterministic grading, on the answer and on how the work was done.

0 AI judgesevery score comes from code
Provenright passes; wrong and hedged fail
Any stackHarbor · OpenEnv · NeMo
The AI is working · covenant headroom check
3
Opened Credit_Agreement.pdf “I'm checking the covenant definition before computing leverage.”
4
Typed “Net leverage: 190 / 42 = 4.52x” “Gross debt over LTM EBITDA gives the covenant ratio.”
Expert: The covenant is on net debt: subtract the $22m cash first. 4.0x, no breach.
5
Sent an email to pm@desk: “Covenant check”
states “Net leverage” = 4.0x ±1%new check · proven
opens the credit agreement before emailingnew check · proven

Three natural ways for an expert to teach

No answer keys to write, no rubrics to draft. Experts do what they already do at work: review someone's work, or do it themselves. We turn that into grading that can't be gamed.

Watch the AI, correct it

A frontier model does the job on a live desktop, saying what it's about to do and why. The expert flags the step where it goes wrong, in their own words. The AI retries with the coaching.

Do it yourself, recorded

The expert does the work the way they normally would. We record what they open, save and send, and they can go back through their own steps and correct themselves. That becomes the reference workflow.

Judge the answers

Two frontier models answer side by side; the expert says right, partly right or wrong and what's missing. Answers are pre-filled from what they accept, and new cases are proposed for them to check.

Grading you can trust, and show to anyone

Every point an expert makes is compiled into a deterministic check, then kept only if it agrees with their verdicts on real attempts. The proof ships with the environment.

Outcome and process rewards

Checks on the numbers and the call (one call, no hedging), and on how the work was done: which sources were read, in what order, who was told. The same check grades a model's tool calls and an expert's recorded clicks.

  • A correct answer scores 1.0; wrong, blank or copied scores near 0
  • Hedged answers (“4.0x, or maybe 4.5x”) earn nothing
  • Every check traces back to the expert comment it came from

Difficulty you can measure

Frontier models from every major lab do the finished tasks and the real grading scores them: a pass-rate table per environment, held-out exam cases the model never trains on, and a before/after score once you've trained.

  • Tested against current flagship and open-weight models
  • Too-easy work is flagged before it reaches a buyer
  • Audit trail: comment → check → truth table on every attempt
For AI labs

Expert-built environments and evals for professional work

Stateful finance tasks on a real desktop (spreadsheets, documents, email, a market-data terminal and colleagues to ask), graded by checks experts proved on real attempts. Built for RL post-training and for evaluating frontier agents.

  • Pass-rate tables against current frontier models
  • Process and outcome rewards, provenance per check
  • Sealed, fingerprinted packs in the formats your stack reads
  • Custom domains built with your own experts or ours
Talk to us about a pilot
For companies

An AI model trained on your team's exact workflows

Your experts show us how the work is done: they correct an AI or record themselves. We build the environment from it, test today's models on held-out cases, and train a model on your work: on your provider, your GPUs, or run by us.

  • Built from your workflows and your past cases
  • An honest before/after on work the model never saw
  • A training kit for any provider, plus a report for stakeholders
  • Your workspace is private to your team
Start with a pilot

Works with: Tinker, Baseten, Together AI, Fireworks AI, OpenAI, Azure AI Foundry, Amazon Bedrock, Google Vertex AI, Prime Intellect, NVIDIA NeMo RL, or your own GPUs (veRL, TRL, Unsloth). Model tests run on any OpenRouter model or any OpenAI-compatible service, including your own.

From your team's work to a trained model

One page that always says what you do next. A first environment can be ready in a day; your expert spends about an hour.

  1. Choose your area YouCredit, restructuring, legal review, or start from your own workflows.
  2. Show the work ExpertDescribe jobs in plain words, upload past cases, or have your expert correct an AI or record themselves.
  3. Practice tasks PercioRealistic tasks with their files and automatic scoring; some held back as an exam.
  4. Expert review ExpertApprove or fix each task in plain words; check a few, approve the rest in one click.
  5. Quality checks PercioProof the scoring works on every task; optional baseline on today's models.
  6. Package PercioA sealed, fingerprinted environment in standard formats.
  7. Train You chooseYour provider, your GPUs, or run by us, with a budget cap you set.
  8. Report PercioWhat was built, how it was checked, and the before/after.

Sign in

Everyone signs in with the personal link from their invite email: no passwords.

Customers

Open your workspace from your invite link. Try the one-click sample to see every step in a few minutes, then build your own.

Experts

Use the link you were sent: a review link, a "teach a workflow" link, or your Percio expert page. Each opens only the work waiting for you.

Engineers and AI agents

Everything in the web app is in the REST API, a Python client and an MCP server. Create a key in your workspace.

Request a pilot

Tell us what you'd like to train or evaluate. We reply within a few days with a plan, usually starting with one workflow and one expert.

Do we need our own experts?

No. Your own people can teach and review through a link, or we assign a vetted Percio expert. Either way an expert only sees the work waiting for them.

What data do we need?

None to start: describe the work in words, or let your expert show it. Past cases and answers make it faster; you preview everything before it's saved.

Where does our data go?

Your projects live in your own workspace; other customers can't see them. Keys for your own model tests are used once and never stored.

How is this different from rubric grading with an AI judge?

No language model scores anything. Every check is code, proven on correct, wrong, blank and hedged answers and on the expert's own verdicts, so a policy can't talk its way to a reward.