Ouch. I wish I had read this before dismissing cousin_it's citation of Robin's point.

Ok, my nothing was unjustified. Though I will point out that if I hire for a second-tier engineering organization which pays no higher than it has too, then the blue weasels that I hire will probably not be much better than the red weasels. All the blue super-weasels will get jobs elsewhere. In fact, it will be found that they are not weasels at all, but rather martins or minks.

True, under the assumption that the weasels are selected only by the threshold test. Actually, since time immemorial red weasels program in Fortran and C++ is thought to be a blue weasel domain. Therefore few red weasels actually plan to be hired as a C++ programmer, only those who are extraordinarily apt apply for such a job. As a coincidence, among the weasels who apply the average C++ ability is significantly higher withing the red subset.

Depends on the distribution of C++ ability. Suppose C++ weasels are a mix of normal weasels and geniuses, that geniuses have far higher ability than normal weasels, and that across both groups blue weasels are on average better by a constant considerably smaller than that difference. Your test could leave you with mostly genius red weasels and a mix of normal and genius blue weasels such that the average ability of red weasels who pass is higher.

Alternatively if far fewer red weasels learn C++ and the red weasels who do are selected for aptitude the average aptitude of red weasels who learn C++ could be higher than that of blue weasels.

Thanks! You could produce Perplexed's claimed outcome by fiat: use your magic detector to hire weasels so that they fit the desired distribution :-) Or you could set the threshold higher for red weasels and get the same result. Both options seem unsatisfactory...