Model cascade

A cascade is a ladder of models ordered by cost. Every request starts on the bottom rung. If the answer passes a check, you return it; if not, you retry on the next model up. Unlike a classifier router, a cascade doesn't predict difficulty in advance. It finds out by trying.

Everything depends on the check. Good escalation signals are cheap and verifiable: output that fails schema validation, a label outside the allowed set, a refusal, a missing citation, an explicit "unsure" field in the output, or a rule your code can test. Asking a judge model whether the answer is good also works, but it adds cost and latency to every request, including the ones that pass.

The economics are easy to get wrong. Escalated requests pay for both calls and take longer, so a cascade saves money only when most traffic stops at the first rung and the check reliably catches the failures that matter. Too lenient, and bad cheap answers slip through; too strict, and you pay twice on most requests. Track the escalation rate, the blended cost per request, and the quality of answers that didn't escalate, since that's where silent failures hide.