The methodology for this seems to be asking the model for a pairwise preference between each possible pairing. So there’s two questions for each pair, to account for ordering effects.
But they don’t multi-sample?? doesn’t seem like good science to me https://x.com/CAIS/status/2067283549909287169