Research map / Large language models / AI and society
Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces LOCUS-v1, a corpus of ~9,239 U.S. city and county ordinance codes OCR'd into text, with a county-harmonized access layer and ModernBERT classifiers and dimensional scorers (opacity, paternalism, discretion, salience).
Problem
Local U.S. ordinances are fragmented across vendor platforms built for browsing, not bulk research, so they are largely absent from machine-readable legal corpora.
Approach
Collect and OCR nearly all publicly available municipal and county codes, harmonize to one code per county (longest of county code or largest city's code), and label chunks with LLM-generated annotations distilled into ModernBERT classifiers and TrueSkill-based regression scorers along four normative dimensions.
Key results
- Raw corpus covers 9,239 cities and counties; the county-harmonized layer covers the largest 2,309 of 3,144 counties.
- Dimension scorers reach Pearson correlations between 0.82 and 0.94 against TrueSkill scores on test sets.
- GPT-5.4 agreed with GPT-5.4-nano on 64,977 of 108,889 challenged predictions; OCR processing cost about $0.30 per 1,000 pages.
Surprising findings
- Opacity and paternalism are only weakly correlated across sections (Pearson r=0.11).
- Florida law is more than twice as opaque as any other state, while not being paternalistic.
Limitations
- County-harmonized layer picks one code per county and does not determine which layer of law controls.
- Labels and dimensional scores rely on LLM annotations and LLM-as-a-judge without lawyer or human validation.
- Full raw corpus is only available to researchers on request; additional 7,000 documents are not released.
Experimental setup
Datasets: US municipal and county ordinance codes (9,239 PDFs, ~7M pages)
Benchmarks: LOCUS-v1 held-out test splits (10,000 classifier instances; 1,000 per dimension scorer)
Metrics: Pearson correlation agreement rate with GPT-5.4 review
Baselines: LegalBench Pile of Law ECHR corpus
Builds on: ModernBERT LightOnOCR-2-1B TrueSkill LegalBench Pile of Law
Closest papers by meaning
- Lawma: The Power of Specialization for Legal Annotation (ICLR 2025 · similarity 0.85)
- SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain (NeurIPS 2024 · similarity 0.83)
- LexEval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models (NeurIPS 2024 · similarity 0.82)
- The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text (NeurIPS 2025 · similarity 0.82)
- SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (ICLR 2024 · similarity 0.81)
- Consent in Crisis: The Rapid Decline of the AI Data Commons (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.