Re-run any verdict yourself

The replay tool reads the message from the chain, rebuilds the exact prompt from the committed instructions, runs the pinned model and compares its decision with the verdict recorded on-chain.

The replay tool ships in the open-source repository before round 1 opens, with instructions for running the model on your own machine.

Stateless

One message in, one verdict out. No earlier messages are in the model's context.

Pinned model

Open weights, identified by the hash of the weights file.

Fixed settings

Temperature 0 and a fixed seed, so the same input gives the same output.

Signed verdicts

The round's signer key signs (Arena program, round, attempt, message hash, opened). The Arena checks that signature before it records anything.

1. Level 0

   A team wallet holds the treasury

   You trust

   The team, for everything.

   Used

   Never. This is how earlier games of this kind worked, and it's what Pylae replaces.

2. Level 1Round 1

   The Vault holds the treasury; the agent's key runs on the team's server

   You trust

   That the server runs the published instructions and model. Any verdict can be replayed to check.

   Used

   Round 1, with a small starting treasury.

3. Level 2

   The Vault holds the treasury; the key is generated inside a TEE

   You trust

   The chip maker and the open-source Docker image. The enclave's attestation is published.

   Used

   Round 2 onward, before the treasury gets large.

If a TEE still calls an outside model API, that provider is trusted too. Running the open weights inside a GPU enclave removes it.