Why did you opt for recovered accuracy as your metric? If I understand correctly, a random probe would achieve 100% recovered accuracy. Are you certain your metric doesn't bias towards ineffective probes?