A one-tailed test is a promise: “Reality will move this way, not that way.” Promises should cost something. If your theory, preregistration, or prior evidence can’t justify a direction, you haven’t earned the efficiency of one tail. Use two and pay the price. But if your model truly points one way (and you’d publish a null if the effect flipped) why dilute power? Write the directional claim before seeing the data. Name the observation that would make you retract it. That small ritual converts preference into hypothesis and keeps you from hiding indecision inside a conservative test. Direction isn’t bravado; it’s accountability.
