Smarter people than me can do it better, but imagine a simple a/b test. 10 videos, 5 with settings A 5 with settings B, give all a [statistics] amount of views. One set gets more slop flags than the other. Tweak settings/model and repeat until they don't get flagged anymore.