Known gaps
What Rounds does not do, where it is weak, and what a pilot on a real clinic group would have to add. Every figure here comes from the seed 23 hold-out run named in the README.
Where the detectors are weakest
- Duplicate records: 67% precision (18 findings, 12 planted). Genuine namesakes (same name, birth date and sex, registered recently) look like a second registration. The operator never merges on its own: a merge is tier 2, and the records lead turns namesakes down when phone and postcode both differ. The cost is the records lead's time on six false findings in the week.
- Misfiled results: 67% precision, 60% recall (9 findings, 10 planted). The detector looks for a hand-filed result whose label matches another patient's open order. It misses a misfiling when the label names the wrong clinic, because the real patient's open order is elsewhere. It also raises correct hand filings where an outside request was filed to the right patient while a namesake waits for the same test; the lab lead turns those down.
- Runs of late cancellations: 80% precision, 86% recall (10 findings, 7 planted). Ordinary late cancellations sometimes cluster at one clinic. Offering freed slots to the waitlist is harmless when the cluster was ordinary, which is why this action is tier 0.
What the world does not simulate
- Nothing reacts to the operator. Patients do not answer offers, reminders or consent requests: an offer stays "offered". A slot offered to the waitlist is not taken. The board shows the operator's action, not its outcome.
- The future is fixed. If the operator moves a booking into a free slot, a booking the recorded diary makes later is not moved out of the way. To keep the diary consistent, the operator only uses slots the recorded diary leaves free, so it sometimes finds fewer free slots than a real diary would offer.
- Times are simple. Results can come back at night, and clinicians acknowledge within the rule regardless of their own hours. Rooms have no equipment; a treatment room is the only special room.
- Undo is seldom exercised. The scripted clinic lead undoes a move into the treatment room, and that only happens when every consulting room is busy. On seeds 7, 11 and 23 no action was undone, so the undo path is proven by tests, not by the recorded runs.
The model
- The headline numbers were produced by rules alone; no model was configured for those runs. The model path (OpenAI-compatible, with fallbacks and a per-run cache) is covered by tests with a stubbed endpoint, not by a scored run. A run with a model is one flag (
--model <name>) and an endpoint in the environment. - Where a model is asked, it chooses between actions the runbook already allows. It never sets a tier, never sees the answer key and never acts on anything outside the policy.
The interface
- One region at a time. The day-board draws one region's clinics (four to eight columns) so blocks stay readable at 1440 pixels; the region strip shows where actions happened, and the rail lists the whole group.
- Blocks sit in their hour, not their minute. Within an hour row, bookings are ordered by start time but not placed at their exact minute.
- Play mode is a single clock. All visitors of one server share the play clock. Live play runs the operator in the server process on the database it was started with; a second live run needs a fresh copy of the world.
- The sign-off queue reflects the end of the recorded run, not the play clock. The action record and the board follow the clock.
Measuring on real data
- Precision and recall here are scores against planted problems. A real clinic group has no answer key: a sample of findings has to be labelled by the people who own each definition, and recall can only be estimated from problems found by other means (complaints, audits, incident reports).
- The world's rates (no-shows, late cancellations, acknowledgement times) are realistic but invented. Thresholds tuned on seeds 7 and 11 will need retuning on a group's own data, by the definitions' owners.
Before anything runs for real
- A read-only reader that maps the clinic system's appointments, results, referrals and messages onto the tables Rounds reads, through a read-only role on a replica.
- Adapters that make the calls the record names (
PATCH /Appointment/…,POST /Communication,POST /Flag), with the clinic system's own authentication and audit. - The group's own rule values, set in the dictionary by their owners, and the policy signed by the people it names.
- A period in shadow mode (
ROUNDS_SHADOW=1): every tier 0 and 1 action is written as "would do", and clinic leads review a sample before any action touches a patient's booking. - Patient messages in the group's own wording and languages, approved by the people responsible for patient communication.