For example, say that 90% of your potential emplyees are crud. You have a position with 100 candidates and 10 are fit for it. Now, you apply a (very well correlated) identifier, it reduces the candiadtes to 50, keeping 8. You still have too many candidates, so you apply another filter, reducing them to 25 in total, with 6 good ones... You can see where this is going, with some luck you'll have: 4/12, 3/6, 2/3, 1/1; you can expect quite often to have: 3/12, 2/6, 1/3, 0/1; with some bad luck you'll have: 3/12, 1/6, 0/3, and no good candidate will make it even to the last round of selections.

Also, there might be a few heuristics you can use to determine the crud-ness of an employer, but they're not necessarily bulletproof.

Example: if an employer uses Clojure/Haskell currently I'd surmise they're not a complete victim of industry's hype cycle. Whereas if the job ad requested node.js ninjas, not so much. I'd guess it'd be easier to introduce node.js in the former, rather than a functional lang at the latter. But this isn't some amazing inferential leap, really. The subtext of the job post is critical.