Why it matters: the buyer may already have a cheaper viable model for some traffic.
A strong answer should cover: routable task features, failure severity, threshold calibration, double-processing, latency and degradation when fallback is unavailable. RouteLLM is a research reference; the local system needs its own test.
Follow-up: “How do we measure the requests wrongly sent to the cheap path?” Ask for the distribution of bad outcomes, not just the percentage of requests handled cheaply.
Interview status: Not yet interviewed.