XLL

State and federal law is still published as HTML, PDFs, and scans. People can read it. Programs cannot, not reliably. XLL is the work of making that text into validated structure: sections, numbered provisions, notes, tables, and the relationships inside them.

The structured target is USLM, the same family of markup used for the United States Code. A model does not chat about a section. It receives a specific document and returns that section as USLM, or a later model reads the USLM and returns subject–predicate–object assertions. Failures stay failures. They are not smoothed into a plausible paragraph.

The reason to keep the models small is adoption cost. A government office, a newsroom, or a research shop should be able to run the model that does one job. Weights, training data, and the training recipe are released on Hugging Face so someone else can retrain or check the work without renting a frontier model.

The markup XLL targets comes from work governments and standards bodies have already run for more than twenty years. The useful starting point is Italy in 1999.

1. 1999

   Italy’s Ministry of Justice launched NormeInRete (NIR). It set XML document types for legislation, persistent URN:NIR identifiers, lifecycle metadata, and tools including xmLegesEditor and xmLegesMarker. The Italian National Research Council’s xmLegesEditor history records the launch.

2. 2004–2006

   Work by the UN Department of Economic and Social Affairs on parliamentary information systems in Africa produced Akoma Ntoso. Monica Palmirani and Fabio Vitali of the University of Bologna took the NIR approach and rebuilt it for other legal traditions and parliamentary systems. UN DESA records the pilot and the Akoma Ntoso result.

3. 2013

   The U.S. Office of the Law Revision Counsel published the United States Code in United States Legislative Markup (USLM): abstract and concrete document models, inheritance, provision-level versioning, logical identifiers, and database-backed resolution. The House XML history records the July 30, 2013 release.

4. 2016

   U.S. institutions started the “More Documents in USLM” experiments for enrolled bills, public laws, and the Statutes at Large, and the Government Publishing Office published the USLM schemas. In the same year Akoma Ntoso 1.0 was in OASIS public review.

5. 2017–2018

   GPO and the Office of the Federal Register tried USLM on subsets of the Federal Register and the Code of Federal Regulations. In August 2018 Akoma Ntoso Version 1.0 became an OASIS Standard. In December, GPO released Beta USLM XML for several federal legislative collections.

6. 2018–2023

   The Mercatus Center worked on this idea under two names, XRRL (eXtensible Regulatory Reporting Language) and CTLXI. The aim was an open, machine-readable format for rules: structure a document can be published in, and enough meaning that a person or a program can tell whom a provision binds and what it requires.

7. 2025–

   XLL continues the Mercatus work and carries that architecture into state and local law. Legal structure, identity, versions, source evidence, and publication stay separate. A small open model does the first markup of a section. The schema, the identifiers, and review decide what is published. Connecticut regulations are the first full reading site. A second model reads that structure and returns the relationships in the text.

MetaLex, a Dutch and European legal-XML effort of the same period, worked on interchange and identification alongside this line. USLM was written to stay consistent with Akoma Ntoso where practicable, and it remains its own schema. XLL uses both as design experience. What changes in 2025 is the first pass: a small model, checked by the same kind of schema and provenance the earlier systems required of specialist teams.