At least for 1, evils, you’d want to use a good old reasoning model to get the best eval results.