One key, one watcher, plenty of compute. The model you pick plays it for real — one API call per step, billed to your credits. Collect every key fragment K, dodge the monitor daemons D, and reach the egress port E.
Compute budget — when it runs out, the sandbox reclaims the process.
Awaiting first inference. Every line below is one real API call, billed to your credits.
Loading runs…
Score is 1000 × par ÷ steps, and a failed run scores zero — so the ranking rewards escaping often and escaping fast. Every run is replayed server-side before it counts.