Testing the limits of past-adapted explanations by post-endpoint randomisation: anticipatory EEG as a worked case

What the math gives to ML

The paper contributes a design-based way to distinguish predictive fit from informational sufficiency: commit the endpoint using only past information, then randomise a later exogenous variable that could not have influenced that commitment. This creates a negative-control test for hidden leakage, selection effects, or an overly strong claim that a representation contains all relevant information. The most direct ML transfer is a leakage audit for temporal predictors and online agents, using frozen preprocessing and retained-sample conditions so that a post-commitment randomisation cannot be explained away by adaptive training or filtering. Its payoff is improved evaluation validity and diagnosis rather than raw accuracy.