For bounded, known-shape tasks, a cheap gate plus deterministic retrieval can beat a language model on cost, speed, and safety — with zero generated tokens. The principle is simple: route first, generate last. Only reach for generation when nothing cheaper can answer the question.
This is an open-ended series of small, honest experiments — single-GPU reproductions, modest in scope and candid about their limits — each one chasing the same question from a new angle. Chapters are added as the work grows.