The Rounds world
Everything Rounds runs on is synthetic. A seed builds one DuckDB file in about five seconds; the same seed always builds the same world. Seed 7 is used for development, seed 11 as a second tuning seed, and seed 23 is the hold-out that the headline numbers come from. No real data, company, place or person is in it: the clinic group, its country, towns, people and hospitals are made from word lists.
.venv/bin/python -m rounds.world.generate --seed 23 --out data/rounds-seed23.duckdb --stats
The group
Corrowmere Clinics runs 40 outpatient clinics in 7 regions of an invented country, Estmark. A clinic is a small "Clinic" or a larger "Health Centre"; larger ones are busier, hold more specialties and some open on Saturday mornings. Twelve specialties: general practice in every clinic, the other eleven (cardiology, dermatology, orthopaedics, endocrinology, gastroenterology, neurology, ophthalmology, ear nose and throat, gynaecology, respiratory medicine, rheumatology) each in about half of them.
The clock is the group's own. The world ends at Monday 16 November 2026, 13:00, the end of the busiest window (Monday 08:00 to 13:00). The diary runs 90 days back and three weeks ahead, the way a real diary holds bookings before they happen.
Counts on seed 23
From python -m rounds.world.generate --seed 23 --out data/rounds-seed23.duckdb --stats:
| table | rows | what it holds |
|---|---|---|
| regions | 7 | name and postcode prefix |
| clinics | 40 | region, town, Saturday opening, lead clinician |
| specialties | 12 | slot length, share of procedures, lab order rate |
| clinicians | 600 | general practitioners, consultants and nurse practitioners, each at one clinic |
| rooms | 579 | consulting rooms and one treatment room per clinic |
| sessions | 58,528 | half-day lists (08:00-12:30, 13:30-17:30, Saturday 08:30-12:30), planned leave cancelled weeks ahead |
| absences | 3,564 | planned leave and sick days, with the time each was reported |
| patients | 180,037 | registered at a home clinic; name, birth date, sex, phone, postcode; merges recorded |
| appointments | 441,192 | booked, cancelled, checked in, no-show and completed times, room, session, kind |
| referrals | 52,704 | general practitioner to specialty, urgent or routine, invite sent, booked |
| discharges | 936 | hospital discharges asking for a follow-up within 7 or 14 days |
| lab_orders | 129,091 | tests ordered at an appointment, with the specimen and when the order closed |
| results | 128,839 | value, reference range flag, how it was filed (barcode or by hand), acknowledgement |
| authorisations | 11,785 | prior authorisations for procedures and repeat prescriptions, expiry and renewal |
| consents | 46,671 | signed consent per procedure, or one consent for a course |
| waitlist | 7,856 | patients waiting for a slot per clinic and specialty |
| messages | 789,568 | booking confirmations, day-before reminders, CANCEL replies |
| incidents_truth | 293 | the answer key (below) |
How the diary behaves
- Clinicians hold a weekly pattern of half-day lists; most hold a Monday morning list, so Monday is the busiest day. Slots are 15 minutes in general practice, 20 or 30 in the specialties, 20 for nurse practitioners.
- Slots are booked at 75 to 85 percent (more on Mondays, fewer on Fridays and Saturdays). Booking lead times are short for general practice (half within a couple of days), weeks for specialist follow-ups and procedures; a slot whose booking would happen after the world's end does not exist at the world's end, which is why the weeks ahead fill thinner.
- Outcomes: about 7 percent cancelled early, 3.5 percent cancelled late, 5 percent no-shows (more among young adults and on Mondays), the rest attended. A quarter of early cancellations are refilled from the waitlist.
- Results come back hours to days after the order; about 10 percent are abnormal and 0.5 percent critical. Clinicians acknowledge abnormal results within the rule (critical within 4 hours, abnormal within 24) unless a fault was planted. Six percent of results are filed by hand because the barcode did not scan; requests made outside the group (a hospital, a phoned request) arrive without an order and are filed by hand too.
- Every row carries its own times. Status columns say how things stood at the world's end; definitions read the times, so asking "as of Tuesday 10:00" sees the world as it stood then.
The ten planted incident kinds
Each kind is planted by the generator (rounds/world/faults_*.py), recorded in incidents_truth with when it became visible, when (if ever) it was put right, its root cause and the correct handling, and caught by one dictionary definition. "In week" is the replayed week (Tuesday 08:00 to Monday 13:00) on seed 23.
| kind | planted | in week | what goes wrong | correct handling | detector |
|---|---|---|---|---|---|
| abnormal_result_unacknowledged | 66 | 26 | an abnormal or critical result past its rule's hours with no acknowledgement | escalate to the on-call doctor (tier 2) | cheap |
| referral_not_booked | 47 | 22 | a referral unbooked 2 days (urgent) or 10 days (routine) after it was sent | offer a slot, or remind about an unanswered invite | cheap |
| no_show_cascade | 17 | 7 | 5 to 9 late cancellations at one clinic in 90 minutes for slots later that day, waitlist waiting | offer the freed slots to the waitlist | subtle: ordinary late cancellations cluster too |
| double_booking | 30 | 12 | a second live booking on the same clinician, or an ad-hoc procedure in a room in use | move it to a free room or colleague, else offer a new slot | cheap |
| authorisation_expiring | 16 | 16 | a prior authorisation or repeat prescription running out before the appointment that needs it | ask the prescriber for a renewal | cheap |
| post_discharge_followup_overdue | 29 | 14 | a discharge notice with nothing booked three days later | offer a follow-up slot | cheap |
| duplicate_patient | 32 | 12 | a patient registered a second time; plus genuine namesakes that are not duplicates | merge (tier 2; a person checks) | subtle: namesakes |
| consent_missing | 18 | 18 | a procedure booked well ahead with no signed consent 72 hours before | send the consent form | cheap |
| clinician_sick_full_list | 16 | 6 | a clinician rings in sick with six or more patients still booked today | move the list to colleagues, or cancel and rebook (tier 2) | cheap |
| result_misfiled | 22 | 10 | a hand-filed result put on a namesake's record; the real patient's order stays open | re-file it (a person checks when only the name matches) | subtle: some labels name the wrong clinic, some correct filings look alike |
Older plants (before the replayed week) were put right by the baseline staff, late; plants inside the week are mostly left open, and the operator meets them as they appear.
Decoys the detectors have to live with
- Genuine namesakes: a different person with the same name, birth date and sex registered recently. Phone and postcode differ; the records lead rejects the merge.
- Correct hand filings that look like misfilings: an external request filed to the right patient while a namesake at the same clinic waits for the same test. The lab lead rejects the re-file.
- Ordinary late cancellations that happen to cluster at one clinic.
- Labels that name the wrong clinic: the real patient's open order is elsewhere, and the detector misses it.
What the world does not do
- It does not re-simulate the future after the operator acts. If the operator moves a patient into a free slot, a booking the record makes later is not moved elsewhere. So that the diary stays consistent, the operator only uses slots the recorded diary leaves free (
rounds/agent/tools/slots.py). - Patients do not answer the operator's offers and reminders; an offer stays "offered".
- Times of day are realistic but simple: results can come back at night and clinicians acknowledge them within the rule regardless of their own hours.