I just wanted to affirm what you are doing here.

A key point here is that P-Values optimize for detection of effects if you do everything right, which is not common as you point out.

> Thompson/multi-armed bandit optimizes for outcome over the duration of the test.

Exactly.