A new proof by Ng, Li, Gu and Thompson shows that for constructed stream-classification tasks, quantum memory stays bounded where any exact classical watcher needs memory that grows without limit, and that for each classical agent below the threshold there exists a suitable input distribution under which its performance approaches guessing. We read the result as the mathematical skeleton of machine learning for ISR, weigh what it would let an edge platform notice on a fixed power budget, and walk the honest distance between a first proof and anything a program office could field.