The same way we currently train any personalization engine using a "maximize revenue" setup – but allow the revenue to be reduced by negative events (even to below zero).Cancel flows only have one reliably measurable metric for training – and it's the thing we don't want them to do. I can't train it...