An infrastructure layer that sits between a request and an AI model, deciding what actually needs to happen: whether the answer can be resolved deterministically from governed knowledge, whether a smaller model can handle it with constrained evidence, or whether a frontier model call is genuinely required.