Seems to me a more reasonable threshold would be at most 10^-9 probability of the conclusion being false (in the Bayesian sense of course), with the prior chosen by the editor and reviewers (and the models being selected upon also being agreeable to them).