Saying It the Law’s Way

This chapter builds the first of the pipeline’s three model jobs: a small, fast model that restates the learner’s question in the words a vehicle code uses, so that retrieval can find the law.

On the deployed pipeline, in early October 2026, what is the highest posted speed limit was refused in ten of the fifty-one states. The states’ codes set their limits in sections that say “maximum speed limit”, and the laws that say “posted speed limit” are the golf-cart and scooter rules, which contain the question’s words and do not answer it. By words, and by meaning too, those rules came first. Texas’s 70, 75 and 85 mph laws ranked 25th, 45th and 46th, outside anything the model would read.

Reworded the way the statute puts it, “maximum speed limit on a highway in miles per hour”, the same three laws ranked 1st, 2nd and 4th.

Nothing in Chapters 4 and 5 can do that rewording. Both rank by the question as asked, and both are right that the golf-cart rule is the closest match to the question’s words. What is needed is a translation from the learner’s vocabulary to the law’s, and that is a language task, which makes it a job for a model: a small one, asked one thing, with a strict shape for its reply and a short leash.

To rephrase a question, in this pipeline, is to produce one or two searches in the code’s vocabulary that would find the law answering it. A search is a short phrase that retrieval runs exactly as it runs the question: ranked by words, ranked by meaning, and fused.