Clinic dictation runs on a model that never dials out

Inside that container, a speech model and a small LLM run as daemon programs on localhost. Exec transcribes, summarizes into the chart format, writes the result. The container's network is set to block at the host, so nothing inside can reach the internet even if a dependency tries. No transcription API, no LLM provider, no vendor on the audio's path.

Hoody's API sees that the container exists and when it was created. It can't see the audio, the transcript, or the model's output; none of it traverses a Hoody tier.

The transcript and the model that made it share a disk, and that disk is in the clinic's rack.