Contribute

Turn research you have already done into a benchmark task

Earn 12 points, become a co-author

A merged task you authored earns 4, one you reviewed earns 1. 3 authored tasks gets you there, as does any mix that adds up. Points land on merge.

Who should contribute

What makes an ideal task

All three. A task that misses any one will not merge.

Your own work

Real research you carried out, not a problem invented for the benchmark.

Did I do this myself?

Weeks of effort

At least two weeks of genuine effort, with or without an agent helping.

Did it really take weeks to finish?

Verifiable

The result is right or wrong, and a script can tell which.

Can a script grade it?

One excellent task is worth more than many mediocre ones.

The four steps

  1. 01

    Ideate

    Pick a project that meets all three. Bring it to group chat or confirm with a maintainer before you build.

  2. 02

    Create

    Write the task package: prompt and metadata, Docker environment, mentor skills, oracle, verifier.

  3. 03

    Test

    Run the oracle, then at least one agent with and without skills, over multiple trials.

  4. 04

    Submit

    Fork the repository and open a pull request against main.

What you must write yourself

Use an AI assistant for the plumbing. The science has to be yours.

Never delegate

  • The prompt body — written by hand, in imperative prose
  • The oracle solution, deriving the answer by computation
  • The scientific judgement about what counts as correct
  • The claim that this reflects real research practice

Fine to delegate

  • Dockerfile scaffolding and pinning dependencies
  • Boilerplate for the verifier test harness
  • Formatting metadata and frontmatter
  • Tidying prose you have already written

The final submission

Three things.

  1. 01

    A PR from your fork

    Fork this repository and open a pull request against main here. One task per PR, touching only files under tasks/<task-id>/.

  2. 02

    A detailed PR description

    The history of the task: what the original work was, its start and end dates, and how many working hours you spent. Plus the physics it exercises and where the data came from.

  3. 03

    A local test results report

    What you ran and what happened: the oracle at reward 1.0, and agent pass rates with and without skills across multiple trials. Example task: PR #2.

Before you open it, both of these have to pass, with the oracle returning reward 1.0:

bench tasks check tasks/<task-id>
bench eval run --tasks-dir tasks/<task-id> --agent oracle --sandbox docker

Give real dates and an honest hour count, and report only the runs you actually completed.

Ten minutes of triage can save a weekend

Bring your idea to Discord before you build it.