Every prediction is graded against reality once it matures (4 h):
true positive = breach predicted and it happened ·
false positive = predicted but did not happen ·
false negative = breach happened without an amber warning ·
true negative = correctly quiet. These verdicts feed the
self-calibration loop automatically.
Alert episodes (amber/breach → back to normal)
| Site | Sensor | Started | Ended | Duration |
Peak risk | Breached? | Resolution | Acknowledged |
Verified prediction outcomes (positives & misses first)
| Site | Sensor | Predicted at | Predicted 4h |
Actual 4h | Error | Risk % | Expected crossing | Verdict |