Ordinary JavaScript. Edit, save and run on the same server.

Save & run

One coding agent edits the scene or program, runs it, reads the actual result and repairs it. Uses your local Codex sign-in.

Start agent

Everything being built: work in progress, commits, the catalogue and the latest recorded runs. Refreshes every 10 seconds while open.

WorldBench: each agent writes one program per task in its own workspace; the task's checker grades it on the public variant and on hidden variants. Latest attempt per agent. Click a variant to replay the graded run.

Compare the selected recording with an earlier run of the same task and scenario.

Reference recording

Measurement	Reference	Selected

One scene renders at a time. A faster failed run earns no improvement credit. Compute time is informational; production time comes from the simulation.

Scenario coverage

Scenario	First try	Selected code	Retries

First attempts include failures. Selected code requires the same program, scene and shared skills across every scenario. Versions of the work order, checker and engine must match. Retries and code versions measure repair effort; they do not imply a count of human interventions.