Model Router.
Not every job needs the biggest AI model. A small, fast decision model called Jev looks at each task and says which model should do it. Plain code makes the final call, using rules I set.
Pick a task.
These numbers are written by hand to show how the judgments are shaped. They are not live Jev output.
Jev judges. Code decides.
Jev doesn't write anything. It can only answer three kinds of question: pick one from a list, yes or no, or a score. For each task it answers a few of them at once: which model fits, whether the task touches anyone's personal data, and whether I should see the result before it goes out.
The rules live in ordinary code, where I can read them and change them. If Jev isn't sure, the task moves up to a stronger model. If it touches student data, it stays inside tools I've approved. If it needs me, it waits for me.
The payoff is simple. Routine work runs fast and cheap, careful work gets the best tools, and the decisions that matter stay with a person.