Reducing the usefulness cost of the blocking protocol can simultaneously improve usefulness and safety, because the evasion gradient comes from the reward difference between intervened and non-intervened trajectories, and a protocol that rarely intervenes generates almost no signal that RL can reinforce.