Pay for exposure, not evaluation.

You are charged for the identities your experiments actually expose. Checking a flag locally costs nothing, however often your code does it.