Until it's known that the same queries will return the same responses every time (if that's what's supposed to be happening), there's no point in looking at what operators are doing. Some part of this has to be controlled in order for any type of experiment to yield useful results.
on the other hand I've also had projects where the database was redundant and queries were 'ad hoc' and 'known' to not match but the data was not judgment based and so was more likely to be relatively reliable. In these cases I started at the query end...
it all depends.
you're right, it's process error