A county procurement agent uses a large proprietary LLM to reason about high-stakes vendor-compliance determinations, but routes routine vendor FAQ questions to a smaller, cheaper model. What is the underlying selection principle?

Tiering models by task stakes lets an architect reserve the most capable - and most expensive - reasoning for decisions where getting it wrong carries real consequences, like a vendor-compliance determination that could affect a contract award, while routing lower-stakes, well-bounded questions to a smaller model that answers just as well for less cost and lower latency. This is a deliberate architectural pattern, not an accident of which model happened to be handy. Running the large model for everything and treating the small model as a mere outage fallback misreads the pattern as reactive rather than a proactive cost-and-risk design choice. Assuming a newer model automatically outperforms an older one on every task ignores that recency isn't a proxy for fitness - a smaller, well-scoped model can outperform a larger general model on a narrow, repetitive task like FAQ answering. And assuming an SLM is better suited to compliance work because it's domain-specific inverts the actual tradeoff: SLMs are chosen for speed and cost on simpler tasks, not because they out-reason a larger model on complex, high-stakes judgment calls. The scope caveat: stakes-based routing needs a reliable way to classify incoming requests by risk, or a genuinely high-stakes question can slip through to the smaller model. A concrete check: audit a sample of FAQ-routed requests to confirm none of them were actually compliance-sensitive questions in disguise.