Model Selection Policy

Where switching is expensive, organizations standardize on one model and defend the standard. Where switching is cheap, they can express a policy instead — bulk summarization to an inexpensive model, nuanced reasoning to a strong one, anything touching regulated categories to local inference. The policy is more valuable than any single selection, because it survives the market moving underneath it and it makes the reasoning explicit rather than historical.

Centralpoint selects per request across OpenAI, Anthropic, Google Gemini, Microsoft Copilot and embedded Llama, Qwen and ONNX, so a selection policy is expressible as configuration rather than as an architectural commitment. Because classification is a property of records, locality can follow sensitivity — regulated categories routed to local inference while general work uses whichever provider is most economical.