Intelligent routers can route effectively between local models and SaaS solutions like OpenAI. They decide based on context like cost, but cost can be the wrong axis for data governance. "What is the expected bonus this year for John Doe when he has a salary of $100,000?" is structurally trivial, so it routes externally to a cheap model, and the data is gone before any policy engine sees it. Cost-optimized routing is, by default, a data exfiltration path. This session presents a layered trust barrier from three components with separated responsibilities: llm-d-sc classifies prompts on two orthogonal axes complexity and sensitivity on CPU in milliseconds. Praxis is the proxy where policy lives, applying the routing matrix as the product of both axes and owning the egress gate utilizing guardrails like NVIDIA Nemo guardrails. I will demo how these technologies work together to secure semantic routing in a hybrid or multi-cloud configuration with a multi-tiered approach.