Yeah, I agree with that. It's not even particularly surprising from a scientific perspective: any scientist knows that if you measure A, B, C, and D, you don't necessarily know anything about union(A, B) or union(A, B, D), unless you also measure pairwise and higher-order interactions. At some point, instead of doing more A/B tests than there are atoms in the universe (since this explodes exponentially), it might be worth thinking about the problem using human intelligence.