NMR staking gives prediction submissions economic weight. It also makes the meta-model’s concentration depend on funding choices, outside capital and allocation settings. A concentration statistic describes those weights; it does not by itself count independent ideas.
Revised September 8, 2026. Claims that 1/HHI measures independent models, or that payouts inevitably eliminate weak models and compound winners, are withdrawn.
A weight is not a complete measure of belief
A participant willing to risk more NMR may be expressing confidence, but it may also have more capital or different funding constraints. Constant and compound allocation modes change how settled gains affect future positions. A model slot can also change its predictions without changing its identifier.
The Models page reports top-model shares and concentration. A retrospective chart of today’s largest models is labeled as that cohort’s history, rather than the history of whoever led each past round.
Effective N is a concentration equivalent
If stake shares are w, the Herfindahl index is sum(w²) and its reciprocal is the number of equally weighted models with the same concentration. Two perfectly correlated models can lower this index without adding independent prediction information. Account-level concentration needs historical ownership, and prediction diversification needs predictions or defensible dependence estimates.
Atomic principal needs its own interpretation
Multiplying every model’s round stake by the same constant leaves its share and 1/HHI unchanged. Applying a fixed 1 NMR eligibility threshold does not: it changes the included population. That is why fractional positions matter even when concentration formulas stay the same.
MMC60 evaluates residual contribution under a defined reference, target and horizon. Fund portfolio construction remains another layer. I would require prediction-level evidence before translating a score-history correlation or stake concentration into a claim about independent strategies.
Method: the cutover figure uses observed Classic model lists captured September 8, 2026. Concentration relationships above are algebraic; no fund-performance or causal selection model is estimated. See the metric guide.