💻 computer science

Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States

This paper introduces LOCUS, a comprehensive, machine-readable corpus of nearly all publicly available U.S. municipal and county ordinances derived via OCR from fragmented sources, accompanied by harmonized access layers and specialized AI models to enable large-scale legal research and analysis of local laws.

Original authors: Denis Peskoff, Joe Barrow, Christopher Vu, Diag Davenport

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Denis Peskoff, Joe Barrow, Christopher Vu, Diag Davenport

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the law as a massive, multi-layered library. For years, legal AI (computer programs that understand law) has been very good at reading the "big books" on the top shelves: federal laws, state statutes, and court cases. These are the rules that apply to everyone in the country or a whole state.

But there is a whole basement level of the library that has been locked away and filled with dust: local laws. These are the rules your city or county makes about things like zoning (where you can build a house), noise levels, business licenses, and animal control.

The paper introduces LOCUS (Local Ordinance Corpus for the United States), which is essentially a massive project to unlock that basement, clean up the books, and put them on a digital shelf where computers can finally read them.

Here is a breakdown of what they did, using simple analogies:

1. The Problem: A Library with No Catalog

Before LOCUS, finding local laws was like trying to find a specific recipe in a library where every town had its own cookbook, but:

The books were in different languages (formats).
Some were handwritten, some were typed, and some were just photocopies of photocopies (scanned PDFs).
There was no master list telling you which town had which book.
You couldn't just ask a librarian (a search engine) for "all the rules about parking in America" because the books were scattered across thousands of different websites designed for humans to browse, not for computers to scan.

2. The Solution: The "LOCUS" Project

The authors built a giant pipeline to fix this. Think of it as a team of robots with super-powered eyes and brains:

The Collectors: They sent robots to visit thousands of city and county websites to download the law books.
The Translators (OCR): Since many laws were just pictures of text (scanned PDFs), they used a special AI (called LightOnOCR) to "read" the pictures and turn them into digital text, like converting a handwritten letter into a Word document.
The Organizers: They cleaned up the text, removing page numbers and headers, and chopped the massive books into individual "laws" (like separating a single recipe from a whole cookbook).
The Labelers: They used AI to tag every single law with labels like "This is about buildings," "This is about business," or "This is about noise."

3. The Result: A "County-Harmonized" Map

The final product is a dataset covering 9,239 cities and counties. However, to make it useful for research, they created a simplified version called the "County-Harmonized Access Layer."

The Analogy: Imagine you want to study the weather across the US. You don't need a separate report for every single street corner; you need one reliable report for every county.

For every US county, LOCUS picks the "biggest" local law book available (either the county's own book or the book of the biggest city inside that county).
This doesn't mean it solves every legal question (like "Can I build a shed in this specific backyard?"), but it gives researchers a solid, comparable foundation to study local laws across the whole country.

4. Giving the Laws a "Personality Score"

One of the coolest parts of the paper is that they didn't just organize the text; they gave the laws "scores" on four specific personality traits. They trained AI to read a law and guess how it feels on a scale:

Opacity (Confusing vs. Clear): Is this law written in plain English, or is it a confusing maze of jargon?
Paternalism (Protecting You vs. Protecting Others): Is the law trying to stop you from hurting yourself (like a curfew for kids), or is it trying to stop you from hurting others (like a noise ordinance)?
Enforcement Discretion (Strict vs. Flexible): Does the law give police officers a lot of room to decide what to do, or is it a strict "yes/no" rule?
Problem Salience (Urgent vs. Minor): Does the law treat the issue as a huge emergency, or a minor annoyance?

Example: A law saying "No one under 16 can go to a festival without an adult" gets a high Paternalism score (it's protecting the kid) and a low Opacity score (it's very clear).

5. Why This Matters

The authors found that local laws aren't just a random mess. They have a hidden structure.

Cities tend to focus on "nuisance" (noise, pets, public order).
Counties tend to focus on "zoning" (land use, building permits).
Different parts of the country have different "flavors" of laws (e.g., Florida laws tend to be very confusing/opaque, while other states are clearer).

The Bottom Line

The paper claims that by "freeing" these local laws from their dusty, fragmented formats and turning them into a clean, searchable dataset, they have built the infrastructure for the next generation of legal AI.

They aren't claiming that AI can now act as a judge or give legal advice. Instead, they are saying: "We have finally built the map and the library. Now, researchers can start asking big questions about how local laws work, how they differ across the country, and how they affect people's daily lives."

Technical Summary: Freeing the Law with LOCUS

Problem Statement
Progress in legal AI has largely focused on federal statutes, case law, and contracts, leaving a critical layer of American law largely absent from machine-readable corpora: local ordinances. While local codes govern essential everyday domains such as zoning, housing, business licensing, public health, and animal control, they remain fragmented across commercial vendor platforms designed for human browsing rather than bulk research. These platforms lack a central registry, utilize inconsistent navigation structures, and often present content as dynamically generated PDFs or scanned documents. Consequently, there is no systematic, national-scale access to local law, hindering retrieval, comparative policy analysis, and the development of legal reasoning benchmarks that require navigating overlapping layers of jurisdiction (state, county, and municipal).

Methodology
The authors introduce LOCUS (Local Ordinance Corpus for the United States), a comprehensive corpus and a county-harmonized access layer. The construction pipeline involves four primary stages:

Data Collection: The team collected raw data from 9,239 cities and counties, totaling approximately 7 million pages across 9,000+ PDFs. They employed browser automation and vendor-specific logic to extract codes, addressing challenges such as server-side PDF assembly limits, filename collisions, anti-bot measures, and multi-county consolidated cities.
OCR and Preprocessing: To handle diverse document layouts (scanned, born-digital, single/double-column), the pipeline utilizes LightOnOCR-2-1B, a 1-billion parameter vision-language model based on Qwen-3. This model converts page images into Markdown. Subsequent post-processing strips artifacts (headers, footers), merges content crossing page boundaries, and segments the text into individual laws based on section headers.
Annotation and Classification:
- Substantivity Filtering: A zero-shot approach using GPT-5.4-nano identifies and removes structural content (e.g., tables of contents) to retain only substantive laws.
- Classification: Three classifiers (trained on ModernBERT-base) categorize laws by Function (Rules, Enforcement, Structural, Context, Process), Topic (Buildings, Business, Zoning, Nuisance, Other), and Substantivity.
Dimensional Scoring: To analyze legal characteristics beyond simple categorization, the authors trained regression models to score ordinances on four continuous dimensions: Enforcement Discretion, Opacity, Paternalism, and Problem Salience. These scores were generated via a "LLM-as-a-Judge" pairwise comparison method (200,000 comparisons using GPT-5.4-nano) and calibrated using the TrueSkill Bayesian rating system. The resulting scores were normalized to z-scores to train the final ModernBERT regressors.
Harmonization Strategy: To create a usable access layer, LOCUS adopts a transparent simplification: for each of the 3,144 U.S. counties, it selects the single most substantial local code (either the county ordinance or the largest municipality's code) based on document length. This covers 2,309 counties, representing the majority of the U.S. population.

Key Contributions

LOCUS-v1 Corpus: A release containing 2,211,516 text chunks derived from nearly all publicly available municipal and county codes, covering 9,239 jurisdictions.
County-Harmonized Access Layer: A reproducible geographic substrate covering the largest 2,309 counties, enabling national-scale search and comparison.
Structured Metadata: The corpus includes function and topic labels for every segment, as well as continuous scores for opacity, paternalism, enforcement discretion, and problem salience.
Model Suite: A collection of ModernBERT-based classifiers and scorers released alongside the dataset to facilitate downstream analysis of local law.
Processing Pipeline: An open methodology for converting fragmented, non-machine-readable local laws into structured Markdown using OCR and LLM-based annotation.

Results and Analysis

Coverage: The raw corpus spans 9,239 cities and counties. The harmonized layer covers 2,309 counties, accounting for the majority of the U.S. population.
Content Distribution: The majority of the corpus consists of substantive laws (Rules and Enforcement), primarily covering buildings, business licensing, zoning, and nuisance.
Model Performance: The trained BERT-based scorers achieved high correlation with the TrueSkill ground truth, with Pearson correlation coefficients ranging from 0.82 to 0.94 across the four dimensions.
Empirical Findings:
- Opacity vs. Paternalism: These dimensions are only weakly correlated ( $r=0.11$ ), indicating they capture distinct legal characteristics.
- Jurisdictional Differences: County codes contain substantially more zoning material, while city codes contain more nuisance and public-order regulation.
- Regional Variation: Florida laws were found to be significantly more opaque than other states. In the Northeast, counties appear less zoning-heavy and more enforcement-oriented compared to the national average, reflecting different institutional histories.
- Structural Patterns: Local codes exhibit a recurring documentary architecture (e.g., general provisions followed by business regulation, then zoning), suggesting that position within a code carries semantic weight.

Significance and Claims
The paper positions LOCUS not as a final theory of local legal authority, but as infrastructure for legal retrieval, regulatory extraction, and comparative policy analysis. The authors argue that "freeing the law" is not merely a problem of access but of representation. By making local law observable at a national scale, LOCUS transforms a fragmented body of public authority into a structured object for computational analysis.

The authors emphasize that while the harmonized layer simplifies jurisdictional complexity, it provides a necessary substrate for future benchmarks that test a system's ability to reason about overlapping authorities (state vs. local, city vs. county). The dataset enables the study of legal dimensions (like opacity and paternalism) at a scale previously impossible, moving legal AI from simple text retrieval toward the reasoning required to navigate the layered nature of American law. The paper concludes that recognizing the structural and functional variations in local law is essential for building reliable legal AI systems.