On-device AI control plane: compute stays local, governance stays in the cloud

模型进了设备，治理留在云端：端侧 AI 的控制面现状

Microsoft Paint's local AI generation hits the cloud twice: first for prompt review and issuing a serial number plus watermark ID, then again to sign the output with a C2PA credential. Reverse engineering shows watermark injection is a hard gate—failure aborts the image. All six major vendors keep governance in the cloud even when inference runs locally. Regulations only require detectability, not per-user traceability; the extra step is vendors building their own risk controls. Three interfaces reveal the real posture: does the prompt leave the device, who issues the identifier, and how long are records kept. Microsoft has not disclosed retention periods.

Why it matters: A reverse-engineering piece that surfaces concrete control-plane details of Microsoft's on-device AI. Specific engineering facts, numbers, and behavioral contrasts (Paint vs Photos app) hit all three HKR axes. Not scored higher because it's a single reverse-engineering report ...