The numbers, and their limits.
Tournament t001, 1500 held-out simulated outages on a 20 km line, thresholds frozen on validation. The bar and the result are on the proof page.
Minutes held, median, correlated loss
See the tables below.
Bound coverage
See the tables below.
Smallest leak still seen at 30 minutes
See the tables below.
Leak recall inside the window
See the tables below.
Same cases, same rules
Scored against, untouched
- Hold-last-value: today's display
- Suspend-and-resume monitoring
- Open-loop line model
- Wavepin's engine with validity masks
- A learned model with validity flags
Thresholds set on validation data and frozen for every test, including the cases the engine never saw. The truth model differs from the engine's own.
The bar, set before the first run
Correlated loss must change holding against random loss at the same rate, beyond the interval.
Correlated loss held 0.91× as long as random loss at the same rate (90 % interval 0.83–0.98). Which gauges go dark together matters, so the dependency map stays.
First proof
Three weeks: 1–3 h scenarios on Wavepin's 20 km line, a hand-declared map (two radio sites, one span, two UPS), four loss patterns, 2,000 scenarios; minutes to tolerance per pattern against hold-last-value.
How long the estimate holds
| Minutes held, median | Error on minutes, median | Inside 90 % bound | “Valid” but wrong | |
|---|---|---|---|---|
| Holdover: filter + jump risk | 71.0 | 41.8 | 85 % | 4 % |
| Hold last value (today's display) | 62.8 | 68.7 | 65 % | 23 % |
| Suspend, resume when data returns | 62.8 | 15.2 | 36 % | 8 % |
| Open-loop line model | 62.8 | 10.3 | 36 % | 13 % |
Outages with at least one gauge dark, no leak. Minutes held: until the worst segment's true error stays above tolerance for a minute. Error and bound: predicted against true minutes. “Valid” but wrong: checkpoints where the method claimed its estimate was good and it was not.
Leaks inside the window
| Leaks caught | False alarms | |
|---|---|---|
| Holdover: filter + jump risk | 91 % | 6 % |
| Hold last value (today's display) | 71 % | 5 % |
| Suspend, resume when data returns | 51 % | 6 % |
| Open-loop line model | 71 % | 5 % |
| Wavepin's engine with validity masks | 67 % | 7 % |
| Learned model with validity flags | 85 % | 7 % |
| Moving-horizon with bias knots | 66 % | 7 % |
Eight-minute excerpts inside an outage, Wavepin's format. Alarm thresholds set at 5 % false alarms on validation, then frozen.
The floor: smallest leak still seen
| 0–2 min dark | 2–10 min dark | 10–22 min dark | 22–45 min dark | 45–200 min dark | |
|---|---|---|---|---|---|
| Holdover: filter + jump risk | 1 % | 2 % | 1 % | 2 % | 1 % |
Smallest leak size class (share of flow) caught at least 9 times in 10, by minutes dark at leak onset. “Not reached”: no size class got there.
Against the bar: correlated loss against random loss
| Loss rate | Correlated, median min | Random, median min | Ratio | |
|---|---|---|---|---|
| 0–30 % | 66.8 | 79.3 | 0.84 | |
| 30–70 % | 72.0 | 72.9 | 0.99 |
Correlated loss held 0.91× as long as random loss at the same rate (90 % interval 0.83–0.98). Which gauges go dark together matters, so the dependency map stays.
What is proven. What is not.
Proven, in simulation
- Leaks starting inside an outage caught 91 % of the time at 6 % false alarms; a learned model 85 %, today's display 71 %, Wavepin's engine with validity masks 67 % (simulation)
- When it says the estimate still holds, it is wrong 3.7 % of the time; hold-last-value 22.5 %
- The estimate held 71 minutes (median) through an outage; the frozen display 63
- The 90 % bound on minutes held covered 85 % of outcomes
- Correlated loss shortens holding to 0.91× random loss at the same rate (interval 0.83–0.98), beyond the interval, narrowly
Not proven
- When an estimate does fail, the minutes forecast is off by 42 minutes (median); a five-minute open-loop rule is closer there
- The correlated-loss effect is small (9 %); another line or map may erase it
- No real outage log used
- No regulator has accepted a computed blind time
- Satellite latency modelled, not measured
- Two sites failing at once may read as one
The first real number comes from your data.
A month of your exports, replayed offline, gives the figure that matters on real operations. We will show it to you whether it is good or bad.