How temporal discounting causes us to misjudge the implications of late-stage software testing.