DAG sounds interesting. Might help me to solve my biggest challenge with evals right now, which is testing subjective metrics e.g. “is this a good email”