Selective escalation

Selective escalation is a confidence-gated cascade with an accept path and an escalate path: decisions above the threshold are taken as final, and the rest go to a larger model, a trained classifier, or a human. One pre-registered study found a frozen gate beat random escalation but missed its 10% accepted-set error target, and a 10-30% escalation band routed to a local classifier matched or beat the better of the two systems.

The economics are the point: routing to an LLM instead of escalating everything matched that LLM at 24-77% lower API cost in the same study. The design work is the threshold and the acceptance test — a gate calibrated on one task can miss its target on another, so evaluate the escalated band on held-out data before trusting it.