Model selection should be driven by task risk and acceptance criteria, not perceived intelligence alone. A useful routing policy records the task class, latency budget, expected output schema, and fallback. Then the team can compare accepted quality and total retry cost rather than choosing models by anecdote.