Last updated: October 8, 2026
This page describes how the LawDiver database was assembled, which published licenses cover the tools used to read and embed that material, and why the opinions, statutes, regulations, case names, and citations in the database carry no third-party research license. Quotations below are from the publisher's own page. Each one links to the source. A description of the build is also at Sources and Methods.
No Westlaw product was used in any form in the creation of the database. No Lexis product was used in any form in the creation of the database. The opinions, statutes, regulations, citation graph, case names, summaries, and embeddings were not taken from either service, and neither service's headnotes, Key Numbers, Shepard's, or editorial enhancements are in the database.
Every embedding in the database was computed with Qwen3-Embedding-8B (Qwen/Qwen3-Embedding-8B), called through OpenRouter as qwen/qwen3-embedding-8b. The model is an 8-billion-parameter text-embedding model from Alibaba Cloud's Qwen team. LawDiver sends opinion text, analysis cards, and statutory text to that model and stores the vectors it returns. LawDiver does not redistribute the model weights.
Alibaba released the model under the Apache License, Version 2.0. The Qwen team's announcement states: "We have open-sourced this series of text embedding and reranking models under the Apache 2.0 license." The model card on Hugging Face is marked license: apache-2.0, and the license file attached to the weights carries this notice:
Copyright 2024 Alibaba Cloud
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
The permission in that license is the grant in section 2:
Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Redistribution of the model itself, which LawDiver does not do, is conditioned by section 4: give recipients a copy of the license, mark files you changed, keep the upstream copyright and attribution notices, and keep any NOTICE-file attributions. Section 6 states that the license does not grant permission to use the licensor's trade names, trademarks, or product names, except as required for reasonable and customary use in describing the origin of the work. The remaining sections are the standard Apache definitions, patent grant, contribution rules, warranty disclaimer, and liability limit. They are in the license file linked above.
The embeddings were requested through OpenRouter, so OpenRouter's Terms of Service also apply to that call. OpenRouter does not replace the model license. It points at it. From the terms in effect on this page's date:
You acknowledge that the Service enables you, your Authorized Users, and your customers (to the extent you incorporate the Service into your own products and services) access to Models provided by Model Providers. By accessing or using any Model through the Service, you agree, and will ensure that your Authorized Users and customers agree, to comply with the applicable terms for each Model ("Model Terms"), a list of which is provided here. You are solely responsible for reviewing the Model Terms applicable to each Model before accessing or using that Model and for determining whether the applicable Model Terms allow you and your Authorized Users to access and use the Service, Inputs, and Outputs as you intend.
As between yourself and OpenRouter, these Terms govern the parties' rights and obligations with respect to the Service. The applicable Model Terms govern your, your Authorized Users', and your customers' access to and use of the applicable Model to the extent required by the applicable Model Provider. OpenRouter does not modify, waive, or limit any Model Terms unless expressly stated in writing.
You may provide input into the Services, which may include images, data, text, and other types of work ("Input") and receive an output from the Services based on your Input ("Output", and collectively, the Input and Output are "User Content"). By making any User Content available through the Service, you hereby grant to OpenRouter a non-exclusive, transferable, worldwide, royalty-free license, with the right to sublicense, to use, host, cache, store, reproduce, transmit, publicly display, publicly perform, publish, distribute and modify (for formatting purposes only), your User Content solely in connection with operating and providing the Service and, depending on the permission you grant for certain features, to provide to other users, individuals, and/or organizations. You retain copyright and any other proprietary rights that you may hold in the Input. Your ownership rights in the Output are set forth in the Model Terms for each Model you use.
For this model, those Model Terms are the Apache License 2.0 quoted above. The license OpenRouter receives in that paragraph is limited to operating the service. A separate, broader license in section 6.2 of the same terms applies only if prompt logging is turned on in the account. The vectors stored in the database are the numerical output of running an Apache-2.0 model on the opinion, card, or statute text described in the sections below.
Language-model processing of the corpus — summaries, holdings, questions presented, and the classification passes — was done with one of these models:
Luna 5.6 and Luna 6 are proprietary OpenAI models. gpt-oss-120b is an open-weight OpenAI model. Which of those four paths handled a given opinion depends on the lane and on failover. Every path is one of those four. The same OpenRouter terms quoted in the embeddings section apply to every OpenRouter call, including both Luna models and gpt-oss-120b when OpenRouter serves it. The model-specific terms that can apply on top of that are below.
OpenRouter's terms say ownership of output is set by the Model Terms. For the Luna models, the model provider is OpenAI. OpenAI's Services Agreement, effective January 1, 2026, and covering API use by business and developer customers, says:
4.1. Generally. Customer and Customer's End Users may provide Input and receive Output. As between Customer and OpenAI, to the extent permitted by applicable law, Customer: (a) retains all ownership rights in Input; and (b) owns all Output. OpenAI hereby assigns to Customer all OpenAI's right, title, and interest, if any, in and to Output.
4.2. OpenAI Obligations. OpenAI will only use Customer Content as necessary to provide Customer with the Services, comply with applicable law, enforce the OpenAI Policies, and prevent abuse. OpenAI will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use.
4.4. Similarity of Output. Due to the nature of OpenAI's Services and artificial intelligence generally, Output may not be unique, and other users may receive similar content from OpenAI's services. Responses that are requested by and generated for other users are not considered Customer's Output.
Section 4.3 of the same agreement makes the customer responsible for having the rights to the input and for evaluating the output. The input here is judicial and statutory text, addressed in the source sections below. The summaries, holdings, and questions presented stored on an opinion are output of that processing. The opinion text itself remains the court's text.
OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache License 2.0, the same grant quoted above for Qwen, together with a one-paragraph usage policy:
We aim for our tools to be used safely, responsibly, and democratically, while maximizing your control over how you use them. By using OpenAI gpt-oss-120b and gpt-oss-20b, you agree to comply with all applicable law.
OpenAI's model card describes the weights as "available under the Apache 2.0 license and our gpt-oss usage policy." When OpenRouter serves gpt-oss-120b, the OpenRouter terms above apply as well, and the Model Terms are this Apache license and this usage policy. LawDiver does not redistribute the weights.
When the same model is called from Groq, Groq's Services Agreement applies in addition to the Apache license and the usage policy. Groq treats the model as a third-party offering whose own terms control, and it treats the generated text as the customer's data:
"Output(s)" means responses generated by the AI Model Services to Inputs. Outputs are Customer Data.
4.1 AI Model Terms. Customer and its Authorized Users will comply with the applicable AI Model Terms for each AI Model Service that is used under this Agreement. Customer and its Authorized Users are solely responsible for their use of the AI Model Services and for complying with the AI Model Terms. In the event of any conflict or inconsistency between the terms of this Agreement and the AI Model Terms, the AI Model Terms will control solely with respect to such use of an AI Model Service.
8.1 Intellectual Property Rights and Permissions. Except as expressly stated in this Agreement, this Agreement does not grant either party any rights, implied or otherwise, to the other's content or any of the other's intellectual property. As between the parties, Customer retains all Intellectual Property Rights in Customer Data (including in Inputs and Outputs), any Customer Application, Customer Training Data, and Customer supplied AI Model Service. Groq retains all Intellectual Property Rights in the Cloud Services. Customer obtains only a limited right to access and use the Cloud Services.
Groq's section 4.2 also states that Groq is not permitted to use inputs or outputs to train or fine-tune models unless the customer explicitly grants that permission. The AI Model Terms Groq points to for this model are the Apache 2.0 license and the gpt-oss usage policy quoted above.
The corpus was built originally on the Caselaw Access Project of the Harvard Law School Library. Between 2013 and 2018 the library digitized its collection of hardbound reporters — about 40 million pages — into a machine-readable set of United States decisions. Harvard describes the opinions as public-domain case law. On March 26, 2024, the Library Innovation Lab announced that the earlier limits on that dataset had expired:
The Library Innovation Lab is excited to announce that the original limitations on the data available for the Caselaw Access Project expired this month, and that data can now be fully released without restriction on access or use.
As part of our original collaboration agreement with Ravel Law, Inc. (now part of LexisNexis) for the Caselaw Access Project there had been access limitations on the full text and bulk data available, which have now expired.
Though most government documents are in the public domain, including case law, this scope of United States case law had never before been made easily accessible to the public.
LawDiver used that released dataset. The access limits in the old Ravel agreement had already expired. The text was taken from Harvard's release, at case.law and the project page at the Library Innovation Lab. It was not taken from a Lexis product. Harvard's own history page is where the expired Ravel collaboration is disclosed; that disclosure is quoted here so the record is complete.
Two quarterly bulk dumps from CourtListener, operated by the Free Law Project, were used: the June 30, 2026 dump and the September 30, 2026 dump. CourtListener regenerates those files on the last day of March, June, September, and December. Each file is a snapshot of their database at generation time, not a delta. The copyright statement on the bulk-data page is:
Our bulk data files are free of known copyright restrictions.
The same page says the files are provided "for developers, legal researchers, journalists, and the public." CourtListener's Terms of Service, last modified August 5, 2026, define the "Services" as the website, the REST API and its webhooks, the MCP server, and the accounts and membership features. Those terms include the use and citing rules that follow. They are quoted because they are the restrictions Free Law Project publishes for its data and for information derived from the Services:
You will not use, intentionally or unintentionally, the Services or any information derived therefrom in violation of any applicable international, national, federal, state, or local law.
You acknowledge that while judicial opinions, motions, and other filings are generally in the public domain, other court filings may contain third-party copyrighted works, such as books and articles, that may retain copyright protection.
You understand that Free Law Project is not a consumer reporting agency under the Fair Credit Reporting Act ("FCRA") and therefore you agree that you will not use Courtlistener.com and any information derived therefrom: (1) as a factor in establishing an individual's eligibility for credit, insurance, employment, government benefits, housing, or any other FCRA purpose as specified in 15 U.S.C. § 1681b(a); (2) to generate a "consumer report" as that term is defined under FCRA, 15 U.S.C. § 1681a(d); and (3) in any manner that could result in Free Law becoming subject to FCRA.
Attribute honestly. If you republish or display our data, do not present it in a way that suggests Free Law Project produced, endorsed, or verified an AI-generated analysis of it.
LawDiver does not present CourtListener or the Free Law Project as the author of LawDiver's summaries, holdings, or good-law notes. Those notes are LawDiver's, produced by the models named above. The opinions in a dump are the courts' opinions.
Quarterly dumps are still expected to be used. Their job from here is narrow: they supply cases that the court-site harvesters missed. New opinions are taken from the courts and from GovInfo, as described next. A dump is a backfill for a gap, not the way the database stays current.
Current caselaw is obtained by harvesters that read government court websites, and primarily from GovInfo, the U.S. Government Publishing Office's public repository. No account was required. The pages are published by the courts and by the federal government for the public.
Judicial opinions are not copyrightable. The Supreme Court restated that rule in Georgia v. Public.Resource.Org, Inc., 590 U.S. 255 (2020), the case about the Official Code of Georgia Annotated. The holding is the government edicts doctrine, and the Court applied it first to judges:
Under what has been dubbed the government edicts doctrine, officials empowered to speak with the force of law cannot be the authors of—and therefore cannot copyright—the works they create in the course of their official duties.
The animating principle behind the government edicts doctrine is that no one can own the law.
In Banks v. Manchester, 128 U.S. 244, the Court held that judges could not assert copyright in "whatever work they perform in their capacity as judges"—be it "the opinion or decision, the statement of the case and the syllabus or the head note."
An opinion harvested from a court site, or taken from GovInfo, is that work: the opinion or decision, the statement of the case, the syllabus, and the headnote, written by judges in their official capacity. There is no copyright in it to license, and no vendor permission attaches to reading it from the court's own site.
GovInfo states the same rule for works of the United States Government, which is where the federal opinions, the United States Code, and the Code of Federal Regulations on GovInfo come from:
Title 17, Section 105, United States Code, provides that: Copyright protection under this title is not available for any work of the United States Government, but the United States Government is not precluded from receiving and holding copyrights transferred to it by assignment, bequest, or otherwise.
The intent of the section is to place in the public domain all work of the United States Government, which is defined in 17 U.S.C. § 101 as work prepared by an officer or employee of the United States Government as part of the person's official duties.
By virtue of the foregoing, public documents can generally be reprinted without legal restriction. However, Government publications may contain copyrighted material which was used with permission of the copyright owner. Publication in a Government document does not authorize any use or appropriation of such copyright material without consent of the owner.
Federal judicial opinions and the enacted text of federal statutes and regulations are works of the United States Government. The caveat in the last quoted sentence is about a third party's material that a government publication happens to reprint, such as a copyrighted image GPO licensed for its own website. It is not a restriction on the opinion, the statute, or the regulation.
Federal statutes and the Code of Federal Regulations are obtained from GovInfo. The GovInfo policy quoted above is the use policy for that text: copyright protection is not available for a work of the United States Government, and public documents can generally be reprinted without legal restriction. The statutory basis is 17 U.S.C. § 105.
State statutes are obtained by downloading the files a state publishes or by reading the statute pages on the state's own website. Those pages are official acts of the legislature. Georgia v. Public.Resource.Org holds that legislators, like judges, cannot be the authors of the works they create in the course of their official duties, and it applied that rule to the official annotations of the Georgia code, not only to the enacted statutory text. The Court's words:
Because legislators, like judges, have the authority to make law, it follows that they, too, cannot be "authors." And, as with judges, the doctrine applies to whatever work legislators perform in their capacity as legislators, including explanatory and procedural materials they create in the discharge of their legislative duties.
Because Georgia's annotations are authored by an arm of the legislature in the course of its legislative duties, the government edicts doctrine puts them outside the reach of copyright protection.
Reading a statute from the legislature's site, or from a file the state posts for download, does not take a copyrighted work and does not depend on a license from a legal publisher. The same no-login, no-bypass method described below is how those pages were read.
Most citations in the database are read from the case corpus itself. A harvested or scanned opinion prints the citations of the cases it relies on. Those citations are taken from that opinion. They are part of the judicial work described above, not a separate licensed dataset.
Some citations were obtained by reading pages on the open web. In no instance did a person or an agent log in to a site to collect them. No security measure was bypassed. No credential was used, no paywall was defeated, and no access control was circumvented. The page was already public.
Case names carry no separate copyright exposure. A case name is the style of a public judicial proceeding. Banks v. Manchester, as quoted by the Supreme Court in Georgia v. Public.Resource.Org, holds that judges cannot assert copyright in whatever work they perform in their capacity as judges, including the opinion, the decision, and the statement of the case. The name of the case is part of that public record. It is not supplied by Westlaw or Lexis, and no license is required to store it or to look a case up by it.