Funny how if you kept reading before commenting, they addressed that point specifically
We were cautious to only run after each model’s training cutoff dates for the LLM models. That way we could be sure models couldn’t have memorized market outcomes.