This message was deleted.

The order of evaluation shouldn't matter because the 50/50 split is based on the result of a hash of the user id with an experiment specific salt - thus a user is always either in control or treatment regardless of any other attributes being filtered on. For large enough sets of users this will hold valid, but in case the filter causes a bias in the split it means that there's some correlation between the filter and user id distribution.