One-tailed tests need ~21% less sample than two-tailed for the same power — by quietly assuming the variant ca

One-tailed tests need ~21% less sample than two-tailed for the same power — by quietly assuming the variant can't lose.
The saving is real; the assumption is usually false.
If your variant can plausibly hurt (most can), a one-tailed test blinds you to that downside and inflates false positives in the direction you bet on.
— One-tail ▮▮▮▮ sample, ▮▮ safety
— Two-tail ▮▮▮▮▮▮ sample, ▮▮▮▮▮▮ safety
Use two-tailed by default. One-tailed only when a negative outcome is structurally impossible.
Read the number, not the story. [power calc · α=0.05]