Known gaps
Read this before trusting a number on these screens. Feeder runs on a synthetic distribution network; every figure is a score against the answer key that the generator wrote alongside the world, not a measurement of a real utility.
The data is synthetic
- The world is generated from a seed. Vellmoor, its towns, substations, feeders, transformers, premises, meters, crews and storm are invented. No real network, customer or company is in it. Seed 7 and seed 11 were used to tune the detectors; seed 23 is the hold-out the README's numbers come from.
- The faults are planted. Ten kinds of incident are written into the data on purpose, with their correct handling in
incidents_truth. A real network has kinds this world does not, and the planted ones are cleaner than real ones: a stuck register here reads exactly zero; in the field it may creep. - The physics is simple. Feeders are radial lines with transformers in order along them; an outage takes out one transformer or everything downstream of a point. There is no load flow, no phase imbalance, no protection coordination, and voltage is one end-of-line value per feeder.
- Messages are lost at fixed rates. 86 % of last gasps and 92 % of restores reach the head end. Real loss depends on the mesh, the battery in the meter and the weather, and is worse in exactly the storms that matter.
- The meter data is daily totals. Fifteen-minute interval reads are summarised to a day per meter for the last fourteen days; the detectors that work on consumption see days, not intervals.
- The control room is a simulation too. Tickets from calls, crew dispatch, travel and repair times come from a queue model with assumed rates. The operator's speed is measured against that model, not against a real control room.
Who decides: rules first, the model only for a real choice
The README's numbers come from a rules-only run. Where the policy leaves a real choice (a voltage excursion: send a crew or ask for switching; an overload: inspect or move load), a model can be asked through FEEDER_LLM_BASE_URL; without one the fixed rule decides and the ledger says "rules". A model run is not part of the published figures.
Scoring
- A detection counts if it overlaps a planted fault of its kind on the same subject, within 20 minutes. A feeder-level outage detection counts for the transformers it lists. Twenty minutes is generous for outages that the operator sees within five.
- Recall is "seen", not "fixed". A fault detected and then declined, or queued and rejected, counts as recalled. The share handled without a person is reported separately.
- The answer key is the world's own. Faults the generator did not plant cannot be counted as missed, and a detection of a real but unplanted problem counts as a false alarm.
- Tuning seeds are in-sample. Thresholds were chosen looking at seeds 7 and 11; their figures flatter the definitions. Only seed 23 is reported.
Not handled
- Anything on the transmission side, protection settings, and disconnecting a customer are never the operator's call (tier 3 in
policy.yaml). - Switching is a request to the control engineer; the operator never operates a switch.
- No adapter calls a real OMS, MDM, workforce or SCADA system. The call on each ledger row is the one an adapter would make.
- Crews are 60 interchangeable teams of four kinds; skills, overtime, shift ends and travel on blocked roads are not modelled.
- In shadow mode (
FEEDER_SHADOW=1) there is no answer key; a pilot needs a plan for labelling a sample of the "would do" rows.