Define latency constraints, provider allowlists, model priority and multi-tier fallback chains
Evaluated on every request before model invocation