Preregistration

Preregistration is the discipline of locking a test's design, the population being scored, the cutoff date separating training data from scored data, and the specific criteria that will count as success or failure, and recording that design before any result is generated. Without preregistration, a backtest or a claim-activation evaluation can be quietly adjusted after a preliminary, unfavorable result to find a different framing, a different population cut, a different success threshold, that happens to produce a more favorable outcome, which turns the exercise from a genuine test of the model into a search for a description of the model's output that looks successful after the fact. This platform requires preregistration specifically for the evaluations that activate a realized-outcome prediction claim for a given asset class, because the stakes of that specific claim, and the temptation to find a framing that supports it, are high enough to justify the discipline. A preregistered evaluation's design should be recorded in a form that cannot be altered after results are seen, ideally alongside a sealed baseline documenting the model version being tested, so that a counterparty can independently confirm the design was genuinely fixed in advance rather than described that way only after the fact.