Is the model actually getting dumber, or are we just reading tea leaves from single samples?