One concrete data point for the thread. The boring version of this is the one that works, and the boring version is: measure a baseline, change one variable, wait, measure again under the same conditions. Nobody wants that answer and it is still the answer.
Happy to go further on any of that.