From scored volume to estimated prevented failures.
Goal: hold every high-risk shipment before dispatch. Progress is held shipments against flagged high-risk orders.
Prevented failures are estimated from flagged volume × verification × observed rate. Refunds you process are real; prevention is the estimate — labeled as such.
Avoidable causes first. Each row ends with the concrete fix.
Pseudonymised IDs. Two or more adverse outcomes each.
Merchant verdicts on holds. A rising false-positive rate means the threshold is too tight.
Every automated and manual decision, newest first.
WAPE of past 7-day forecasts vs realized cases. Under 25% is high trust.
Days with at least 3× the trailing median. Worth a look before they become tickets.
What-if math on your observed efficiency. Planning aid, not a prediction.
The retention surface: top risks, accuracy, prevented story. Email delivery is configured at deploy time.
Brier score is the mean squared error of predicted probabilities. Lower is better. Only the per-shop adapter learns; the base model is fixed.