AI Law Lab
A research platform for running reproducible experiments on how language models reason about law — contracts, case law, statutes, regulation, and jurisprudence — executed on the University of Wyoming's own DGX Spark cluster.
Sign in with UW sign-on, Google, Microsoft, or an AI Law Lab account. University of Wyoming accounts are approved automatically; others are approved by an administrator.
Why it exists
AI-generated legal claims are on the rise: they are easy to produce but hard to check. An answer that looks convincing once says little about whether a model reasons soundly, cites real authority, or behaves consistently when the facts shift. AI Law Lab exists to put those claims to the test as experiments: every run records its exact configuration, the exact source facts that it searched, and a full trace of what the model did, so the result is something that can be re-run, compared, cited, and analyzed.
Law is also argued, not only read. Simulated negotiations, mediations, and hearings have long been how lawyers learn to advocate, and legal case role-play brings that practice into the lab. Agents representing each side hold their own objectives, bottom lines, confidential facts, and case files, and must decide what to disclose and when to concede, while a moderator keeps the exchange moving. Running the same dispute many times shows how models argue, persuade, and settle under pressure — whether they hold a position or give it away, cite their evidence honestly, and keep a confidence — and gives students an opponent to practice against and instructors a transcript to teach from.
What it does
Grounded retrieval
Build curated libraries from uploaded files, web pages, and searches of CourtListener, the Federal Register, the eCFR, govinfo, and SEC EDGAR. Every document is chunked, embedded, and anchored to its pages; every change is recorded as a new library version, so a run can be repeated against the library exactly as it searched it. Each library produces a citation list for any version.
Document analysis
Ask a question of your libraries directly: it is broken into sub-questions, each is searched across one or more libraries together, and the answer is written from what is found. Add a contract or opinion and the question is answered from it instead, with the libraries as supporting authority. Each numbered citation is traced to its document, library, and version, and one that names no retrieved passage is flagged as unsupported, not hidden.
Agentic research
Write a research brief; a planner turns it into sub-questions and picks the libraries worth searching, and you review the plan before it runs. Research agents then work on the sub-questions in parallel, searching the libraries and, if you allow it, the legal databases and the open web, for as long as the question needs. A lead agent writes the answer. Everything read is saved to a library first, so every citation leads to a fixed source, and every step is recorded for audit.
Legal case role-play
Stage a negotiation, mediation, or hearing between agents with their own background, temperament, bottom line, confidential facts, and private memory. Give each side its own case files: citing a passage discloses it as an exhibit the other side can answer. A moderator keeps turns fair and steps in when talks stall; an evaluator assesses the outcome and the disclosures. Draft a cast from a news story or opinion, or upload agent files written in plain Markdown.
Capabilities at a glance
How a study runs
- Build your libraries. Upload documents, add web pages, or search the online legal databases and add the results you want. Each library keeps a numbered history of versions and a citation list for each.
- Design an experiment. Pick a mode in the builder — no JSON required — and choose the libraries it draws on. For a role-play, write the cast, upload agent files, or let the AI draft one from a link, file, or pasted text; choose the legal sources everyone shares and give each agent its own case files.
- Launch a run. Fill in the run's inputs, and optionally change its libraries or pick earlier versions of them. The run records exactly what it searched; work is dispatched across the Spark cluster and streamed back live as a trace.
- Read and trace the result. Every citation links to the document it came from; References list each cited source with its library version and every place it was cited, and a role-play shows what each side disclosed. Export the run as a PDF, or run it again with the same library versions to reproduce it.
Runs execute real language-model calls against shared research hardware. Please be considerate of others using the cluster.
Notes
-
AI Law Lab: System Architecture
A one-page diagram of the whole system — from sign-in and the web portal, through the experiment manager and the three experiment modes, to the shared retrieval, tools and memory, the router, the Spark cluster and the record every run leaves — followed by a reference to each tier: how libraries are versioned and searched together, how a citation is traced back to its source, why inference is admitted through a single bounded queue, and what the server fetches from outside.
-
Moderating Multi-Agent Legal Role-Play with a Small Local Language Model
How the role-play moderator chooses speakers, detects and breaks impasses, and decides when a scene is over; what each participant is allowed to see; the research each rule rests on; and what a live 24-turn mediation showed about where it still falls short.
Acknowledgement
AI Law Lab was created with a seed grant Faculty Award from the University of Wyoming's School of Computing to Prof. Kipp A. Coddington (College of Law), in collaboration with Dr. Jian Gong (School of Computing); Michael Killean and M.J Jonnala (UW Advanced Research Computing Center, ARCC); and Emma-Jane Alexander and Jesse A. Ballard (UWIT); as well as numerous external collaborators, advisors, and evaluators.