> Let’s examine a third version: the people investing $2.4M into this book use AI in their evaluation process.

On that note, LLMs exhibit "self-bias", so knowing what kind of model does the judging helps you tailor your thing to its prejudices, even if just by using the same model to "spruce up the woring so that this wins a competition."

Further along that spectrum is adversarial models where you create a new model designed to bamboozle the judge, but that's a lot more work.