Numeraire

On February 20, 2017, Numerai announced one million Numeraire allocated to 12,000 existing data scientists according to past tournament performance. The whitepaper by Richard Craib and collaborators addressed a problem that a leaderboard alone could not solve: a model can fit exposed historical data without generalizing to unseen outcomes. Staking was meant to make confidence costly to fake, not to certify intelligence.

The June 20 Ethereum launch was a separate event. Numerai reported sending more than 1.2 million tokens to 19,000 data scientists and said there would be no ICO. It also announced a move away from Bitcoin payments toward ether and NMR so Ethereum applications could use the proceeds. The distribution rewarded an existing contributor base; the deployed token then let those contributors put that reward behind their next predictions.

On July 8, 2019, Numerai announced NMR 2.0: the maximum fell from 21 million to 11 million, further issuance and token upgrades were disabled, and tournament functions moved to separate contracts. Even the old function name mint became a burn operation. This separated a stable risk asset from a competition whose rules and infrastructure would continue changing.

That distinction survives in the current product. Unstaked models can build a live score history, but the docs assign them zero weight in the Stake-Weighted Meta Model. Staking on one’s own submission adds economic exposure and influence on that combined prediction. It does not turn the participant into a fund investor; Numerai still evaluates predictions and organizes the tournament.

Numerai dates the Classic tournament’s Atomic cutover to round 1343, opening August 28, 2026. A continuous stake could previously support several unresolved rounds through accounting. Atomic staking instead backs each round, staker and model position with its own locked NMR. The Classic configuration describes 64 overlapping rounds: a one-NMR round setting would require 64 NMR once all those positions are live.

The transition is not a single release of every old stake. Earlier rounds settle under the old system while new positions use the new route. The current migration table gives Classic 24 business days, but a retained FAQ says 64. That conflict prevents a confident completion date. For a participant, the relevant distinction is between an unresolved old position, a funded new position and idle NMR that can be withdrawn.

Atomic accounting also changes what a loss looks like. In an automated claim-and-restake flow, a model’s burn can remain deferred debt offset by a later payout from that same model. It is not spendable NMR and cannot fund another stake. A direct claim realizes the outstanding burn instead. A negative score therefore does not always mean canonical supply falls at that instant.

The system still relies on Numerai to authorize submissions, compute scores and publish the settlement commitments that claims prove against. The participant’s practical decision has become more precise: how much NMR to commit to the next round, whether to compound and when to stop opening new positions. Numeraire’s experiment continues at that boundary between believing a model and bearing the cost of its unseen results.