Catch a mistake before it becomes a booking.

Select a node to inspect its records. Replay uses saved evidence; Run live makes new paid model requests and requires operator access.

Chance of a costly mistake

All registered travel cases, grouped into eight equally weighted failure families. Live requests and injected failures are excluded.

Credible intervals are model-based and prior-sensitive, especially for rare events. Conservative upper bounds are reported separately. Unknown proposal losses are never replaced with an invented point value.