Policies

A corpus is only worth what can be trained on it, and a model’s own report of how well it did is worth what any self-reported number is. Every entry here was rolled out by this server, on the same eight starts, in the station’s own dynamics.

A policy minted on the contract is a different object: a cap table over the trajectories that would train a model, so a licence fee splits to everyone who recorded them. Minting one trains nothing.

POST to /api/policy with a submitter address, a label, and the weights — mean, std and three weight matrices with their biases, in the architecture above. Nothing you send is executed: a policy is three thousand numbers and evaluating one is arithmetic, which is why this can be open. An entry is keyed by the hash of its weights, so the same model submitted twice is one entry.