A benchmark is only as good as the model you use to grade it

I built a pytest harness that runs the same set of questions through five language models at once - a...