The method: turning a frontier model into your firm’s associate
The one idea, the four layers, and why the method — not the model — is the proprietary advantage.
Research & methodology
Read the work before you hire anyone.
PLI publishes its methodology and its open-source model research in plain language for lawyers and in full depth for engineers. The GitHub and Hugging Face pages are the source record; these notes explain why the work matters and how it fits together.
Methodology & curriculum
The method PLI teaches.
How a firm turns a frontier model into a reliable associate — the layer of skills, workflows, retrieval, evaluation, and guardrails that is the real thing worth owning. This is also the curriculum behind PLI’s enablement training.
Skills, not chatbots
Why a library of narrow, well-described skills beats one giant prompt — and why the description is a classifier, not decoration.
Deterministic document generation
Treat document automation like a compiler: the model drafts, deterministic code builds, and formatting never drifts.
Retrieval over your own work product
Your archive is the asset. Ground drafts in your own approved language — and never let a style source supply the facts.
Evaluating legal AI without fooling yourself
A single demo proves nothing. Verified model identity, honest provenance, and confidence intervals instead.
Fail-closed legal AI tooling
How legal AI should behave when the evidence is weak: stop, record the reason, and require human review.
Practising private legal AI
Explainers for lawyers.
Plain-English pieces on the questions every firm evaluating AI should be able to answer — written to teach the vocabulary, not to sell.
Why on-prem isn’t paranoia
What “the data never leaves the building” requires, and why owning the stack is now the affordable option.
Hallucinations and sycophancy, in plain English
Why models invent citations, why they are built to agree with you, and the one question to ask any vendor.
Owning your AI stack: the economics
Why today’s cheap frontier pricing has a history, and where owning the stack pays off.
Ethics & risk
Professional responsibility.
The old duties — competence, confidentiality, candor — applied to a new tool. General information, not legal advice.
The confidentiality math
How professional-responsibility guidance treats AI, why the terms of service matter for privilege, and why “the AI said so” is not a defense.
PLI Labs research
Open-model work, in the open.
Source-level research on legal model pruning, Hugging Face releases, and owned-hardware inference — summarized in plain language, linked back to the public record.
Legal REAPs, not finetunes
Why PLI Labs prunes open models around legal capability instead of training on client data.
Qwen3.5 legal REAP release
Two Qwen3.5 legal-compression candidates published for lawyer-supervised evaluation.
What the Qwen3.5 run showed
A candid finding: the model is prune-resistant, so PLI Labs published a conservative cut rather than a bigger headline.
Hugging Face release index
A practical guide to the MiniMax-M2.7 legal REAP repositories and GGUF artifacts.
A 122B-class legal lane on four B70 cards
A reliable four-card layer-split serving lane, benchmark results, and stack guardrails.
Serious legal AI on V100 hardware
Why topology, backend, and KV-cache settings decide whether an older owned GPU fleet works.
B70 and V100 launch profiles
Measured successful launch settings and token rates for owned inference lanes.
Everything PLI Labs just published
The public drop: Labs repo, model cards, Hugging Face pages, and hardware research.
Source record
Public GitHub & Hugging Face.
The current PLI Labs public repository is ProprietaryLegal/pli-labs. The research summaries above link to the underlying artifacts published there and to the matching Hugging Face model repositories.