What LOCUS contains
On August 31 at UC Berkeley, Professor Diag Davenport presented LOCUS, a corpus of 9,239 codes from US cities and counties, totaling approximately seven million pages. [1 · UC Berkeley Law]
The official Berkeley Law event page confirms the presentation and the project’s subject. The corpus makes it possible to study a layer of law that is formally public but remains fragmented across many local sources. [1 · UC Berkeley Law]
Local codes are particularly difficult to process automatically: their structure, archive completeness and publication methods vary. A unified corpus enables systematic comparison and reproducible model evaluation. [1 · UC Berkeley Law]
What LEXCHEX measures
LEXCHEX is a benchmark of leading AI systems’ ability to answer questions grounded in local law. It tests not general legal knowledge but the ability to find an applicable source within a large, heterogeneous collection. [1 · UC Berkeley Law]
The project emphasizes that a published law is not necessarily easy to observe, compare or search by machine. Different formats, versions and publication methods complicate both human work and AI evaluation. [1 · UC Berkeley Law]
The benchmark also sets a stricter standard of usefulness: a system must ground its answer in a specific applicable rule, not merely produce plausible text. This distinguishes confident writing from verifiable legal work. [1 · UC Berkeley Law]
A commercial layer over public documents
Analysis: Similar problems arise in many fragmented domains, including instructions, warranty terms, certificates and regulations. Collecting documents is not enough; product value emerges through normalization, version control, data provenance and answer testing. [1 · UC Berkeley Law]
Trust in an algorithm depends on its ability to show the applicable source, its current version and the limits of the system’s authority. LOCUS and LEXCHEX offer a research model for turning scattered public material into a corpus, a benchmark and an application layer. [1 · UC Berkeley Law]
The next important outputs will be published evaluations of specific systems and descriptions of errors, such as selecting the wrong jurisdiction, using an outdated version or failing to cite a rule. [1 · UC Berkeley Law]
Sources
- UC Berkeley Law — LOCUS / LEXCHEX event — Official event page, August 31, 2026