I think it's an implicit assumption that a big complicated rule will be very picky in what it accepts, meaning hits are rare and will reveal more information than misses. They think all the complications are on the input side, so they're trying to pick interesting inputs instead of interesting outputs.
Come to think of it, I might be making this mistake but for non-boolean functions. I try to always have tests that pass in interesting values, but maybe I should consciously also try to have tests that get back interesting values. For example, when testing a classifier I should make sure all the classifications are possible. Or, when testing some mathematical function implementation, I should have tests with inputs that cause the result to be exactly -1, 0, 1, and 2 (if applicable).
Come to think of it, I might be making this mistake but for non-boolean functions. I try to always have tests that pass in interesting values, but maybe I should consciously also try to have tests that get back interesting values. For example, when testing a classifier I should make sure all the classifications are possible. Or, when testing some mathematical function implementation, I should have tests with inputs that cause the result to be exactly -1, 0, 1, and 2 (if applicable).