Reading a training set's own annotations names the motion it lacks: no benchmark, no evaluation split. The gain tracks which deficient cells were closed, not who closed them. Withhold the diagnosis and the same generator, at the same dose, is harmful more often than not.