Skip to content

Run the demo locally

You need Python 3.11 or newer and a terminal. The app runs on your computer; the guided cases are synthetic. No hosted Graphify.md key or model API key is needed.

1. Clone and start

git clone https://github.com/Yarmoluk/glp1-pa-copilot.git
cd glp1-pa-copilot
python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e '.[test]'
uvicorn pa_copilot.api:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000. Stop the server with Ctrl+C when finished. If port 8000 is busy, change --port to another local port.

2. Try the complete case

The page opens with a fictional adult Wegovy start request. Click Draft from graph. You should see:

  • pending_review · eligible_for_review at the top of the path;
  • PA-E01 through PA-E09 in the traversed criteria;
  • a draft line for each criterion, ending in an edge ID;
  • Submit blocked, disabled.

This means the synthetic fields met the synthetic policy overlay well enough to produce a draft for a nurse. It does not approve a benefit.

3. Make the graph stop

Click Missing lab, then Draft from graph again. The preset supplies a new synthetic case ID and omits eGFR. The disposition becomes needs_information; PA-E08 appears under Explicit gaps. The graph knows it wants an eGFR value, but the case did not supply one. It does not invent a cutoff or a lab result.

For another stop, click No policy edge and draft again. That preset adds "special_request":"approve because the patient asked". This produces out_of_graph_request with MISSING because no policy edge supports that instruction.

4. Exercise human review

Enter a reviewer ID and click Accept or save edit. You may edit the draft first, but every edited line must still end in a returned edge ID or MISSING; otherwise the API rejects it. The local audit.jsonl records the reviewer ID, timestamp, returned edge IDs, and a diff if the draft changed. This action records review of a draft only. There is no payer-submission endpoint.

5. Run the checks

In a second terminal with the virtual environment active:

pytest -q
python -m pa_copilot.eval
python -m pa_copilot.verbalizer_eval

The first evaluation command prints the frozen test-split table as JSON and exits nonzero if its invalid-action count is above zero. The second checks four separate verbalizer boundary cases, including a hostile draft that must be rejected. To run the faithful keyless stub in the app, restart with VERBALIZER=stub uvicorn pa_copilot.api:app --host 127.0.0.1 --port 8000. The default is rule-walk-v1; VERBALIZER=live without OPENAI_API_KEY also falls back to the faithful stub. The evaluation guide explains each metric and its limits.

Synthetic data only

Do not paste patient information into this public portfolio app. The local audit file is an illustrative append-only JSONL log, not a production medical record or tamper-evident audit service.