How we validate
Every figure on this site comes from data the client already had, tested on periods the model never saw. This page is the method — including the machines where it did not work.
Calibrate once, then replay month by month.
- Calibration
- Calibrated on one full year, on data never seen during calibration.
- Replay
- Replayed over 12 rolling windows — retrained each month, tested on the next, each month evaluated once out of sample.
- Baseline
- Every configuration compared with random alerting at the same alert budget, so a result has to beat chance at equal cost.
- Causality
- Only the past is ever used to produce a score, so what is validated offline is what runs in real time.
Two of four machines paid off. We publish the other two.
From the aerospace and defence line. Neither V1 nor V3 is priced into the case.
| Machine | Downtime anticipated | Break-even per false alert | Verdict |
|---|---|---|---|
| V4 | 217 h · 56% | 5.9 h | pays off under any assumption |
| V2 | 47 h · 50% | 37 min | pays off if a simple inspection |
| V1 | 6 h · 16% | ~5 min | not cost-effective |
| V3 | 3 h · 4% | ~6 min | no usable signal in the log |
On V3 the major breakdowns are preceded by no build-up of micro-stops. There is nothing to see in the log, and it is not a tuning problem. We state that in the deliverable rather than tuning until something appears.
Alert load across the line runs at two to four alerts per machine per week, and roughly one alert in six is followed by a serious stop within the hour. That ratio is in the deliverable too, because it is what decides whether the alerts get acted on.
The zone comes first. Then we count the runs inside it.
Not a prediction
The recommended zone was delimited first. Only then did we look at which past production runs fell inside it, and what they actually yielded.
A statistical floor per zone
Each zone carries its own floor. Across the full 336-run history, no run inside a zone ever fell below that zone’s floor.
Within reach of the line
On a representative run, 8 of the 10 retained levers were already inside the recommended range — the zone is not a different factory.
Extrapolation is labelled
Any range that goes beyond what the line has already produced is flagged as extrapolation rather than presented as a finding.
Not every lever is a setpoint
Ambient temperature is observed, not set. It guides when to produce rather than what to adjust, and we say which is which.
Confirmable in a week
One run takes about five hours. Running the next productions inside the zone and comparing yield to history gives a first verdict within a week.
A validated case, or a documented no.
Both are real outcomes. The second one you keep too.
Name one line and one target variable.
We read the export you already produce and tell you within a week whether it carries the levers — before you commit to anything.
