LabPolicy

Train and benchmark

The Lab trains and benchmarks frontier AI models.

An agent learns what its environment rewards. The Lab builds the environments, rewards and benchmarks that teach frontier models the rules before they are deployed.

1,600+
Members in our record
3
Public benchmarks we track
8
Areas of research
379
Publications by members

Benchmarks

The public evidence on what frontier agents can do and where they fail, from the people who measure it best. We build our environments against it.

Agents

How long a task a frontier agent can finish, half the time.

Data fromMETR

Time horizons 15 models

The length of task, measured in the time it takes a human expert, that each frontier model completes with 50% success. The newest model on the chart finishes tasks that take an expert about 17 hours.

Source
  1. Claude Mythos Preview17.4h
  2. Claude Opus 4.612.0h
  3. GPT-5.25.9h
  4. Claude Opus 4.54.9h
  5. GPT-53.4h

What the Lab Builds

The environments programme is in development. Each environment teaches one thing an agent has to get right before it is trusted to work alone.

The Lab

RL environments

  • Sandboxed tasks with real tools, files and permissions.
  • The rules an agent will work under, drawn from the statute or policy itself.
  • Delivered behind a standard interface your training stack can call.
Build an environment

The Lab

Rewards and rubrics

  • Graders written by members in computation and in law.
  • Rewards that pay for keeping to the rules, not only for finishing the task.
  • Every rubric published with the reasoning behind it.
Write rubrics with us

The Lab

Evaluations

  • Tests run before an agent is deployed, and after.
  • Scores against the rules that apply, task by task.
  • Results briefed to the legislators and officials who ask for them.
See the evaluations

The Reward Is the Guardrail

An agent learns what its environment rewards. So we write the rules down, build the world the agent works in, pay for the right behaviour, and check that it held.

  1. 01

    Specify

    Write the rules down.

  2. 02

    Build

    Build the world the agent works in.

  3. 03

    Reward

    Pay for the right behaviour.

  4. 04

    Evaluate

    Check, publish and brief.

Get in Touch

For frontier model providers and research partners. Tell us what your agents do and the rules they have to keep.