Nurse-signed by default.

Companion listens at the bedside and drafts a nursing note. It is a good draft. It is often the best draft a tired nurse will see at the end of a shift. And it does not post to the medical record on its own. A human reads it, edits it, and signs it. That is a deliberate choice, and it is worth saying out loud why.

The auto-post demo is the one that closes the room. The AI hears the visit, writes the note, drops it into the chart, and the screen fills in by itself. It looks like magic. It is also the version that gets a facility cited, a nurse disciplined, and a resident harmed when the model hears denies chest pain and writes denies chest pain over a quiet yes.

The chart is a legal artifact

A nursing note is not a productivity output. It is the record a surveyor reads, the document a family's attorney subpoenas, and the line a licensing board points to when a license is at risk. The signature on it is not decorative. It is the nurse asserting, under their credential, that this is what happened. An AI cannot make that assertion. It does not have a license to lose.

Remove the typing, not the responsibility

Companion's job is to remove the typing — the keyboard time, the late-night catch-up, the carbon-copy paragraphs nurses paste because the shift is ending. The judgment stays with the nurse. chart-by-exception only works as a workflow when the exception is caught by a human who was actually there; otherwise it is a fiction the chart tells about itself.

Concretely, every Companion draft lands in a review queue tied to the assigning nurse, with the source audio and observation timestamps attached. Edit, sign, post. Nothing reaches the EHR until that signature event fires. The next thing we are shipping is a side-by-side diff between the draft and the signed version, so the model learns from each correction the nurse already had to make.