cross_iid() calls score(de, target, condition) – data first, parameters
second, the order nle()'s estimator is trained in. A ratio has no target
and no condition, and reads better with the parameter first, so the flip
happens here rather than in de_log_ratio()'s signature.
