Playground

Audit your agents, right here

The real Python library runs in your browser through WebAssembly. Your data never leaves this tab.

Python starts on your first click or keypress (about 10 MB, cached afterwards).

Load trajectories

JSONL, one attempt per line: {"model","task_id","run","seed","success"}. Extra fields are fine.

Drop a .jsonl file or click to browse

Or try an example:

Edit the raw data

Confidence level Bootstrap resamples

How many tasks and runs per task do you need to detect a difference? After an analysis, the variances are filled in from your data.

Target difference Task-level variance Run variance Cost of a new task Cost of one run Power

The planner appears as soon as Python has started.