Runtime Footprint
The canonical evaluation path is intentionally light. CPU-only models, a few gigabytes of RAM, no GPU required. The replay client runs on a laptop. The CPU-only constraint exists for two reasons: it keeps the participation floor low for verifiers, and it forces the canonical model choice toward open weights that anyone can audit, instead of a frontier API model that a verifier could not reproduce.
| Profile | What it does | Hardware |
|---|---|---|
| Replay client | Verifies state advances against chain logs | Ordinary CPU laptop |
| Coordinator | Serves the API, scores submissions live | CPU server sized for throughput |
The coordinator runs the same code on a larger machine for throughput. The scoring math is identical between coordinator and verifier.