- A transcript that says who spoke
- Each line carries Clinician, Patient or Other. Lines the model could not place stay unlabelled, and a mislabelled exchange can be corrected without retranscribing the visit.
- A note in SOAP form
- Subjective, Objective, Assessment and Plan, as four sections a reviewer accepts or rejects one at a time. A draft is not a note, and nothing in it reaches the chart on its own.
- Codes checked against the catalogue
- Every proposed ICD-10-CM code is looked up in the loaded code catalogue. A code the catalogue does not have is dropped, and the description shown is the codebook’s wording, not the model’s.
- Dropped codes are counted out loud
- When suggestions are dropped, the review screen says how many. “No codes were suggested” and “three were suggested and none of them exist” are different facts about a visit.
- Header codes marked as non-billable
- A category or header code that will not stand on its own on a claim is labelled as such, quietly, on the row. It is a label rather than a warning, and it changes no control.
- Coding guidance from rules, not impressions
- Deterministic guideline checks run over the grounded codes and carry their citation. Rule-backed recommendations are listed ahead of the model’s, and an empty list has to state why it is empty.
- Hints from the clinic’s own care templates
- When a code maps to a condition the clinic has authored care components for, the matching goals and interventions are shown as hints beside the recommendation. Nothing is prefilled and no care plan is created.
- A partial transcript says so
- If part of a recording did not transcribe, the review screen states it in plain numbers before the draft is read. A short transcript is a fact about the note, not an error to be hidden.