A document-QA assistant answers employee questions. On your evaluation set, a fast model matches a larger model on 85% of questions (simple lookups) but is much worse on the 15% that require reconciling several policies. The larger model with extended thinking meets the chat latency target only at the median, not at p95. Which configuration is easiest to justify to stakeholders?

Answer: C — Route by a measured classifier: fast model for lookups, larger model for multi-policy questions, streaming for both

The evaluation shows where the expensive configuration earns its cost, so route: fast path for the 85%, strong path for the hard 15%, with routing accuracy measured and streaming to improve perceived latency in chat. (a) misses the p95 target for all traffic to help 15%. (b) accepts known failures on the high-value cases. (d) is a global compromise the evidence doesn't support, and it still under-serves the hard cases.