But for someone whose whole job is to find, and invest in, "the next Google", you'd think there could be some sort of recognizable features that a Bayesian classifier could learn as positive weights (as, if there weren't, that person would guess correctly at no greater a rate than chance, and would therefore be replaced by a small shell-script/die roll.) Those recognizable features should then be able to be decomposed into discrete features—which could be written down as a checklist.

To put it another way—there are only two ways to do a job whose output is boolean (invest/don't invest): either algorithmically, or randomly. Any judging algorithm can be approximated by a checklist.