Skip to content
wildenv한국어Follow
Blue engraving of a river valley opening to the sea. On a rocky bluff stands a small field station with a weather mast; one person writes in a notebook on a rock while another walks past with a laptop bag.

Team Attention · Research program

Real work, from the wild, made into environments.

wildenv is Team Attention's research program on one question: how do you turn real work, done by real people, into tasks, graders, and environments for reinforcement learning?

We publish the methods, the checks, and what went wrong.

Read the research

We publish methods and findings. Raw data stays private.

Method

Three steps from real work to an environment.

  1. Engraving of a camp desk outdoors with a laptop, a notebook, a small video camera on a tripod and a folded map.

    01

    Start from real work

    • Real agent session logs, plus the files involved.
    • The workflow, explained by the people who do it.
    • Explicit consent and agreed rights before any use.
  2. Engraving of a glass terrarium holding a miniature landscape with ferns, a stream and a hut with a weather mast.

    02

    Build the environment

    Only reviewed, sanitized projections become environments. From them we synthesize tasks, graders, and environments, and we own their quality.

    Coding follows Terminal-Bench, app workflows follow AutomationBench, and knowledge work follows AA-Briefcase.

  3. Engraving of measuring instruments: a balance scale, dividers, a folding ruler, a magnifying glass, a stopwatch and blank tags.

    03

    Measure it on 9 models

    We calibrate each task with gold, no-op, near-miss, alternative-correct, and mutant solutions. Then we run it on 9 models across 3 tiers and audit the failure traces.

Painted blue sky with tall white clouds and a weather mast with a wind vane, a cup anemometer, an instrument shelter and a small solar panel.

Published research

We publish what we learn.Methods, checks, and failures.

Every study is public. It is how we check our own pipeline, and how anyone can check us.

Start with the task checks

Partners own the data. We own the quality.

Partners bring the raw data. They own it and are responsible for its rights. We synthesize the tasks, graders, and environments, and we own their quality.

For companies

Bring a whole workflow. We build its environments.

Email us about your data

The research is how we earn trust. The work that matters most starts from a company's own data: the full record of a workflow, not scattered sessions. The company keeps ownership. We build and audit the tasks, graders, and environments, and report how models do on its work.

  • Ownership of the data stays with the company.
  • Raw data is not exported. We work inside the company's environment or on a de-identified copy, and only tasks and graders come out.
  • Scope, rights, retention, and deletion are agreed in writing before work starts.

Thesis

From workflow evals to a company model, and back to evals.

The moat is not accumulated data. It is the pipeline that keeps producing graders and environments cheaply.

  1. 01Enterprise workflow evals
  2. 02RL data
  3. 03A model specialized for that company
  4. 04Its operating data feeds new evals

Then the loop repeats.

Unit economics · measuring now

We measure human time per graded task.

Our target
≤ 1 hour≈ $100 per graded task
Lab prices
≈ $200–2,000per task
Cost per graded task, log scale

We are measuring this now. We do not report a measured value yet.

session2eval · What we learned

What we learned from our pipeline.

These are internal numbers from synthetic rewrites.

See the 368 checks for synthetic task defects
  1. 01

    We map each session to the format of the nearest authoritative benchmark: Terminal-Bench-style coding, AutomationBench-style app workflows, and AA-Briefcase-style knowledge work.

  2. 02

    We calibrate each task with gold, no-op, near-miss, alternative-correct, and mutant solutions.

  3. 03

    We run each task on 9 models across 3 tiers.

  4. 04

    We audit failure traces. In our first round, every defect that changed a verdict was found in failure traces, not in text review.

  5. 05

    After the fixes, one workflow task went from 2/9 to all strong models passing, while weak models still failed.

FAQ

Questions about wildenv

What is wildenv?

A research program run by Team Attention. We study how to turn real work into tasks, graders, and RL environments, and we publish the methods and findings.

What is an RL environment?

A task for an agent, the environment it works in, and a grader that checks the result. Models learn from them through reinforcement learning, and the same tasks work as evals.

Can I contribute my own sessions?

Yes, as an optional research collaboration. It is unpaid. Leave your email below and pick the agents you use. We will write to you first, and nothing is used without your explicit consent.

What happens to raw data?

Raw data stays private. Only reviewed, sanitized projections become environments, and consent can be revoked.

Can a company work with you on its own data?

Yes. Email bong@team-attention.com. The company keeps ownership of its data, and we agree on scope, rights, and deletion before any work starts.

Who runs wildenv?

Team Attention, an independent research institute in the spirit of Artificial Analysis.

How do I ask for removal or a correction?

Email bong@team-attention.com.

Follow the research

Get new research by email.

Leave your email and we will send new studies. If you want to collaborate with your own sessions, pick the agents you use. Collaboration is optional and unpaid.

For questions, company data, removal, or correction requests, email bong@team-attention.com.

Interested in collaborating? Agents you use Optional