I get fewer matches than I expect with this, but I think it has with
the chaining of \< and \> to do (i.e. the \> eats the \< from the next
substring that would have matched). You could perhaps give up on the
leading word anchor, since peoPLE are unlikELY to write lIKE this (and
if they do, they deserve to be killfiled).
Given that suggestion, how about
[A-Z]+[^A-Za-z0-9]+[A-Z]+[^A-Za-z0-9]+[A-Z]+
i.e., cap(s), non-letter(s), cap(s), non-letter(s), cap(s)
(Too likely to get a false positive, unless you score it so that it
only hits when several lines match.)