Many years ago, I used to work on grants from the Postal Service. The goal was to read addresses and sort mail automatically. The metric we wanted to minimize w

Many years ago, I used to work on grants from the Postal Service. The goal was to read addresses and sort mail automatically. The metric we wanted to minimize was 10*E+R , where "E" is the fraction of errors (where the system returns an incorrect answer), and "R" is the fraction of rejects (where the system doesn't give an answer). The thinking was that it is 10x more expensive to send the letter to a wrong address, than to reject it outright and have it processed by a human. This includes the cost of customer service, anguish, bad reputation, etc., and not just the cost of delivery and re-routing.