Ah, sorry, I misread what the OP meant by "matches" - thought they were referring to all classifier outputs, while they specifically meant the positives. I changed my original comment to better reflect what I meant, even though that makes it a bit of a non-sequitur now.