The former would make any new site inspectors look drunk merely from the low # of sales.
Low number of sales _across_ all sites = just bad image.
Statistics of only rejects of good images would give too much weight to sites with high acceptance rate, so it has to be counter-weighted with 'false-positive' one.
I found different problem - false-positive and false-negative got different weights. After fixing this, numbers (and positions) changed. I'll update original post. I am surprised - one of my worst sellers, FT, appears to be pretty good