Revised September 8, 2026. The original study classified a payout-adjusted change in model stake as a deliberate deposit or withdrawal. That attribution was too strong. Its four behavioral figures and claims about capitulation, profit-taking and deposit timing have been withdrawn.
A residual is not a transaction
The archived study compared 404,219 weekly model-stake observations from July 2024 through June 2026. Subtracting an estimated payout effect from a stake change produces a residual. It does not identify the person responsible, a wallet transfer or a motive.
The calculation also used a uniform 33-day settlement delay. A valid update must use each round's published end and confirmed settlement state. Even with the correct date, a model's stake can change because an account reallocates between models, a model misses a submission, settings change, or capital moves between staking systems. Claim and withdrawal events have their own times.
Atomic migration makes the distinction visible
Classic principal per round fell from about 796,465 NMR to 12,528 NMR at the cutover. That is a change in how capital is allocated to individually funded positions; it cannot be read as a 98% investor withdrawal. Positive-stake model counts fell by about 4%, while a fixed 1 NMR cutoff would incorrectly report many fractional positions as exits.
The current balance view reconciles the legacy and atomic generations at the source timestamp. It answers a different question from the model participation charts, which count observed model slots and their positions.
Measure the decision at the right level
To study economic behavior, keep separate records for external deposits and withdrawals, transfers between model allocations, position commitments, confirmed settlements, claims, idle funds and deferred burn debt. Historical ownership is needed to group model slots into accounts without assigning today's identity to every old observation.
Per-model allocation settings matter too. Reinvestment in compound mode can change later position sizes without a new discretionary deposit. Constant mode can leave released gains idle. Outstanding deferred burn debt belongs to the model and cannot be treated as withdrawable or reusable capital.
Scores still determine rewards under the applicable round policy. The economic commitment is NMR put at risk on a prediction contribution. Neither a stake residual nor a high score by itself reveals why an owner committed that capital.
The website now labels model participation, round principal, settled outcomes and current balances separately. A future behavioral study needs transaction and ownership evidence before it can attribute cash flow or intent.
Sources: Numerai public round-details responses captured September 8, 2026, and the atomic staking specification. The replacement figure uses the same pinned cutover evidence as the round-economics correction.