This is why using other LLMs as scorers for benchmarks and evaluations is such a bad idea, they'll have preferences you can't anticipate and won't understand immediately.