Suppose a factory monitors a 90 MW demand threshold for a given interval tomorrow. The model returns P50 of 78 MW and P90 of 96 MW. If the authorised first action is a low-cost check and an enquiry about flexibility, the range warrants review. P90 estimates a conditional percentile. When calibrated, roughly nine in ten comparable observations should fall at or below it. It does not give an exact exceedance probability at 90 MW.

If a shift review finds that a large start can move out of the same demand interval, present both load distributions and production effects to operations. Deferral can be uneconomic even with risk. To quantify expected cost, estimate exceedance at the threshold or use calibrated scenarios and allow for joint peaks across intervals. State assumptions if interpolating probabilities from a few quantiles.

After the event, record whether an action was available as well as the outcome. An actual of 93 MW exceeds the threshold yet sits below the 96 MW quantile. Save the trigger, check, measured load and authority together. If load was altered, retain the action time and estimated adjustment. The intervened outcome alone cannot score an untouched-load forecast.

Once adopted, forecasts can change the process being predicted. A warning may defer a production start and prevent an exceedance. That is neither automatically a false alarm nor automatically a saving. Preserve the issued forecast, decision log, action interval, actual load and process. Build a reviewed no-intervention range using comparable days or explicit production assumptions where feasible.

Start in shadow mode to verify timeliness and replay performance without requesting new actions. Release alerts gradually to comparable sites or shifts, preserving comparable measurements and operating practices. Review forecast, uptake and action outcome separately. Account for holidays, equipment changes and production shifts through controls or stated ranges rather than attributing every before-and-after change to the model.

Use three measures to guide investment: coverage of actionable peaks, conversion after reviews, and net action value and workload. Frequent misses point to inputs or calibration; alerts without response point to notification, ownership or authority; expensive interventions point to criteria or contract terms. In quarterly review distinguish traceable events, unresolved attribution and changes for the next period.