Small LLM Judges Approved 11% and 41% of Wrong Answers. Then I Fixed My Own Pairwise Test.