Task library
Document environments the way a strong research deck would: clear defaults, precise constraints, and readable benchmark framing.
Browse tasksResearch benchmark platform
Benchmark agents with the calm precision of an editorial research studio.
WorldModel Gym turns environments, uploads, traces, and leaderboards into a public-facing benchmark product that feels intentional from the first click.
Sparse-reward environments with reproducible seeds
Public leaderboard slices across test, train, and continual tracks
Browser upload flow alongside API and CLI submission paths
Benchmark narrative
Build a benchmark surface that communicates as well as the experiment itself.
FastAPI + Postgres + S3Next.js App RouterReal runs live in production
Our benchmark surfaces
Rigorous evaluation, not cherry-picked demo polish.
The site is designed to make benchmark evidence readable: task framing, run uploads, leaderboard slices, and trace inspection all move together as one public story.
Design benchmark tasks with explicit failure modes
Upload metrics, traces, and configs without leaving the product surface
Inspect runs, compare tracks, and share outcomes with one clean URL
Live leaderboards
Compare planning quality, return, and cost in one public surface instead of across notebooks, screenshots, and scattered artifacts.
Open leaderboardsUpload studio
Create a run, attach metrics and traces, and publish it from the browser while keeping automation-friendly CLI and API options.
Publish a runProduct workflow
A single workflow for create, evaluate, upload, and compare.
Workflow Create
Design a benchmark brief that feels intentional from the first click.
Shape sparse-reward tasks, defaults, and success criteria before you ever touch a leaderboard. This is the fastest way to move from a research idea to a benchmark someone else can immediately understand.
Prompt
Frame a partially observable benchmark with delayed reward, reproducible seeds, and a planning budget that matches the story you want the leaderboard to tell.
Step 1
Choose an environment with explicit constraints
Step 2
Set defaults that make the benchmark reproducible
Step 3
Carry the task into evaluation and upload flows
Task framing
Task defaults
Observation mode
Reward design
Task defaults
Observation mode
Reward design
Live benchmark product
Ready to publish planning research without the visual clutter.
Ship new runs, compare them publicly, and use the same benchmark surface in your README, interviews, project portfolio, or research demo.
