know.2nth.ai has fourteen sections and over two hundred pages. Type a question and Jev navigates: at the root it scores every section in one call, in the section it scores every page, and it keeps going until it lands on a leaf or says none of these. Every option's probability is on screen, along with the wall clock and the cost. This is intent routing you can watch.
Each hop is one Choice question over every child of the current page, with each page's description as the option text, plus an explicit "none of these". A second yes/no question asks whether any listed page answers directly. The model only sees titles and descriptions, so it is routing on the index, not reading the leaves. The wall clock includes fetching the index and the network hop to the model. On tasks like this, TypeSafe itself notes the speedup over a chat model is smaller than on pure decisions, because fetches dominate. The point is to watch it decide, not to time it.