What to do next

The corpus is twenty-seven laws. The deployed one is the titles of a state’s code a licence-holder is bound by, and Chapter 9’s chunker is the hundred lines of a four-hundred-line script. Fetching Maryland’s Transportation Article from the General Assembly’s site, cutting it with your chunker, and reading what it gets wrong is a weekend, and every label shape it gets wrong is a line in the deployed chunker with a comment naming the section that taught it.

The eval is twelve cases. Add one every time an answer is wrong, with the figure read off the source and the wrong figure in notContains, and run it before every change to the prompt. The set will be worth more than the prompt within a month.

And the reading copy at https://github.com/AiFirstDevelopment/howards-lantern-rag-pipeline has, beside most of the files quoted in these chapters, the design document that decided each: decisions numbered, acceptance criteria numbered, and in nine of the fourteen a section headed How this fails silently that pairs each one with the test that would catch it. Read the first road-law plan, then the one for the vector index, then the vehicle code, then nationwide, then what happens if, in that order. They are the pipeline’s history, and each says what it tried first and why that was not enough.