Llm Eval Router

Shadow-test local Ollama models against a cloud baseline with a multi-judge ensemble. Automatically promotes models when statistically proven equivalent — reducing API costs with evidence, not hope.