The function
The input is one section as the publisher printed it: the heading, the numbered paragraphs, the notes, the tables, and the links. The output is that same section as USLM, with the hierarchy made explicit and the words unchanged.
Two checks decide whether a section can be published. The XML has to validate against the USLM schema, and the text of the XML, with whitespace removed, has to equal the publisher’s text. A model that rearranges a sentence into cleaner prose has failed, even if the result is readable.
How it was trained
The recipe search started on Qwen 9B. Four machines tried learning rates, epoch counts, and data mixes in parallel, and each candidate was compared with the untouched base model on held-out sections. The point of the small model was to find a recipe before spending a larger training run.
The scaled run is a full fine-tune of Qwen3.8 27B on a mixed set of 5,000 sections: 2,792 gold examples from Government Publishing Office sources and 2,208 silver examples from the Electronic Code of Federal Regulations, chosen to cover titles the gold set does not. Training runs two epochs, completion-only loss, and a 32,768-token cap with no truncation. Sections are the training unit. A section that does not fit is split by a defined rule rather than cut off mid-sentence.
Training loss is not the quality number. Quality is schema validity and source-text fidelity on held-out sections, including tables and long sections. The weights, the training rows, and the recipe belong on Hugging Face with the model card, so the run can be repeated.
Where it stands
Connecticut is the first jurisdiction carried through to a reading site. Sections the model did not get right were repaired against the publisher page and checked again for the same two conditions: valid USLM, and raw text that still matches the source. One section, 46a-5-11, is held back because its printed labels cannot be stored as legal section numbers without changing the text.
Open the Connecticut regulations. Statutes, registers, and the other document types will be separate models and separate pages. They are not extra tasks piled onto this one.