Summary. In a rules-first design, a model helps build a rules base from primary sources, and at question time ordinary code applies the rules. The model's only job is reading the person's words into facts. I tried this for a Minnesota DWI advisor, and two posts report how well it worked and how I tested it.

What the posts say

The pattern, in one sentence.

“use AI to build and maintain a rules base for one body of law, and run the rules, not the AI, when someone asks” — Azure Foundry Agent (Part 6)

Why it is attractive: speed, cost and traceability.

“In this case the answers are cheaper and faster, and each figure traces back to the section of the law it came from.” — Azure Foundry Agent (Part 6)

Where the model sits.

“At question time only the rules run.” — Azure Foundry Agent (Part 6)

A small model is enough, because the job is small.

“Because that job is small, the model can be small: qwen2.5:7b, running locally, with no frontier model anywhere.” — Azure Foundry Agent (Part 6)

Plain rules already read most messages.

“Rules alone, with no model at all, read 73% of 150 fresh messages exactly.” — Azure Foundry Agent (Part 6)

Little is left for the model to get wrong.

“Everything with a legal consequence is code, so there was little for the model to get wrong.” — Azure Foundry Agent (Part 7)

A perfect reader and the real small model scored the same on the 2,000 tests.

“With a perfect reader the advisor also passed all 2,000, and the real small model matched it.” — Azure Foundry Agent (Part 7)

Search engines do not rank a statute above a blog post; the rules base does.

“A search engine ranks pages. It does not rank a statute above a law firm's blog post, and our advisor does.” — Azure Foundry Agent (Part 7)