- Home
- Research
Research & methodology
Read the work before you hire anyone.
We publish our methodology and our open-model research in plain language for lawyers and in full depth for engineers. The GitHub and Hugging Face repositories are the source record; these notes explain why the work matters.
Methodology
The method we teach.
How a legal organization turns a frontier model into a reliable colleague: skills, workflows, retrieval, evaluation, and fail-closed guardrails.
- Start here
The method: turning a frontier model into your firm’s associate
The proprietary value in legal AI is not a secret model — it is the method of using a frontier model: skills, workflows, retrieval, evaluation, and fail-closed guardrails. A practicing litigator's account of the architecture, and how it becomes teachable.
- Skills & routing
Skills, not chatbots
A skill is a compact operating manual that tells an AI model when a class of work applies and what rules to load. Why skills and registered workflows beat one giant prompt for legal work.
- Document assembly
Deterministic document generation
Why serious legal document automation treats the model as a drafter of Markdown and lets deterministic code own formatting, templates, and file construction — producing repeatable, filing-ready output.
- Retrieval
Retrieval over your own work product
Every firm sits on a decade of its own filed work product — its voice, its arguments, its judgment. Grounding AI drafts in that corpus, while never letting style sources supply facts, is the durable edge.
- Evaluation
Evaluating legal AI without fooling yourself
A single impressive demo proves nothing. How to evaluate and select AI models for legal work with real rigor: verified model identity, honest provenance, infrastructure-versus-quality separation, and confidence intervals.
- Guardrails
Fail-closed legal AI tooling
A user-friendly summary of PLI Labs tooling principles and related public legal AI repositories.
For lawyers
Explainers, in plain English.
The questions every legal organization evaluating AI should be able to answer — written to teach the vocabulary, not to sell.
- Risk literacy
Hallucinations and sycophancy, in plain English
Why AI models invent citations, why they are built to agree with you, and the one question to ask any vendor: show me exactly what happens when the model is wrong.
- Deployment
Why on-prem isn’t paranoia
What 'the data never leaves the building' actually requires, how on-premises AI reduces third-party exposure by design rather than by promise, and why owning the stack is now the affordable option.
- Economics
Owning your AI stack: the economics
Why today's cheap frontier-AI pricing looks like earlier subsidized tech platforms, why a firm shouldn't let a core business function depend on rented compute, and where owning the stack pays off.
- Professional responsibility
The confidentiality math
How professional-responsibility guidance treats AI: the duties of competence and confidentiality, why the terms of service matter for privilege, and why 'the AI said so' is not a defense.
PLI Labs
Open-model research, in the open.
Legal model pruning, Hugging Face releases, and owned-hardware inference, summarized and linked to the public record.
- Model pruning
Legal REAPs, not finetunes
A user-friendly summary of PLI Labs legal model pruning work and the MiniMax-M2.7 legal pruning campaign.
- Release
Qwen3.5 legal REAP release
PLI Labs releases Qwen3.5 legal REAP models for lawyer-supervised legal AI evaluation, with Hugging Face models and public research documentation.
- Research note
What the Qwen3.5 run showed
Research notes on PLI Labs' Qwen3.5 legal REAP release, conservative expert pruning, and the Layerdrop6 companion model.
- Releases
Hugging Face release index
A guide to the five public ProprietaryLegal Hugging Face repositories for MiniMax-M2.7 legal REAP candidates and GGUF artifacts.
- Owned hardware
A 122B-class legal lane on four B70 cards
A user-friendly summary of PLI Labs public B70 inference research, four-card llama.cpp SYCL success, benchmark results, and XPU guardrails.
- Owned hardware
Serious legal AI on V100 hardware
A user-friendly summary of PLI Labs public V100 inference research, benchmark findings, and serving lessons.
- Owned hardware
B70 and V100 launch profiles
Successful B70 and V100 owned-hardware launch profiles for legal AI, including public settings, measured token rates, and Hugging Face model links.
- Release index
Everything PLI Labs just published
A user-friendly summary of the June 18, 2026 public PLI Labs GitHub and Hugging Face upload covering legal REAP model cards, B70 research, V100 research, and owned-hardware launch profiles.