CEMI Labs · Legal
Augmentation, on the record.
This is where Lawra experiments with how legal work actually gets done. Everything on this page is a live experiment you can open today — not a roadmap. Each entry says what it is testing, what it has taught us, and where to go and see for yourself.
If we try to make change look like the past or respond to the expectations and structures of the past, it’s not true change, is it?
Carlos Miranda Levy, Coordinator of CEMI’s Enhanced Intelligences
Each entry is a running experiment. The date is when it opened on this site; the proof is the thing itself — open it and use it.
Testing: Whether an adversarial rehearsal — an AI bench and opposing counsel that push back over several rounds — sharpens an argument better than a tool that simply drafts one.
What it taught us: The value sits in the resistance, not the output. The lawyer argues; the machine plays the room and then rules. Three personas review the ruling afterwards rather than a single verdict voice.
Open the simulator →Testing: Whether a question put to several legal traditions at once — Roman, common law, civil code, Islamic, Confucian, contemporary — produces better thinking than a single synthesized answer.
What it taught us: Disagreement is the product. The panel surfaces common ground and productive tensions instead of averaging them away. The seven contemporary voices are openly fictional composites, labelled as such, and none of the panel is ever quoted as a real person.
Convene a panel →Testing: How to produce a coherent 90+ document corpus when no single model call can hold it — a sequential orchestrator that plans sprints, streams each, and lets the work resume after a closed tab.
What it taught us: This became the house template for any agentic tool whose output exceeds one call: plan, stream per step, package as you go, survive interruption. Incorporation and the Founders’ Package both run on it.
Generate a package →Testing: Whether tools that already know your jurisdiction, your citation style and your house playbook beat tools that ask again on every run.
What it taught us: The setup cost is the barrier, not the idea — so the profile is built through a short interview or an uploaded document rather than a form. It is a layer across the suite, not a product beside it.
See what it does →Testing: Whether a catalog organized by the work a lawyer actually does — research and drafting, contracts, litigation, compliance, operations, formation — is more usable than one organized by vendor tier.
What it taught us: Naming the categories after legal work rather than internal pricing tiers changed what people find. Ten of the tools are agentic: multi-step, tool-using, self-directed.
Browse the suite →Testing: Whether short narrated pieces explain a legal-AI tool faster than a page of copy — and whether they can ship without costing a visitor bandwidth they did not ask for.
What it taught us: Twenty-six pieces, click-to-load behind poster frames, so an unplayed clip costs nothing. They are Spanish-audio today; that is the honest limit on how far they currently reach.
Watch the reel →Testing: Whether hard, unsettled questions — AI and copyright, open models, authorship — are better served by taking a position in public than by staying neutral.
What it taught us: Every claim carries a source, and each piece is reviewed in the open by three personas who disagree with each other. Where a holding is qualified, the text says so rather than rounding it off.
Read the corpus →Testing: Whether legal AI can serve nine languages without flattening the legal systems behind them — civil law and common law do not map word for word.
What it taught us: Translation is the easy half. The Spanish and French surfaces are not the English ones restated, and the gaps that remain are tracked openly rather than papered over.
Switch languages →Working conclusions from the experiments above — design positions we hold because of what we built, not research findings. We say which is which.
The tools people come back to are the ones that argue with them — the bench that rules against you, the panel that disagrees, the reviewer that finds the weak citation. Tools that only produce text get used once.
An unsourced legal answer transfers risk to the reader. Returning less, with a traceable source and a stated limit, is worth more than returning more and being unverifiable.
We used to cap reply length by account tier. It produced answers cut off mid-sentence, which reads as a broken product rather than an upgrade prompt. Tiers now govern how many questions you may ask — never how completely any one is answered.
Several personas here are composites, not people. Labelling them plainly costs nothing and is the difference between a useful device and a false credential. None of them is ever given a quote a real person did not say.
The work is open. These are the ways in.
Everything in the ledger is reachable right now — simulate a hearing, convene the council, generate a package. No demo call in between.
Open the suite →Assessment, consulting and hands-on implementation for firms and legal departments putting AI into live matters — with your lawyers, on your work.
Work with us →A problem in practice the lab should test, a tool that should exist, a claim in our writing you think is wrong. Tell us what you would put on the bench.
Write to the lab →Four labs, one conviction: technology earns its place by amplifying people.
Learning, tested live.
Work the floor. Raise the bar.
The impossible, made human.
The Legal Innovation Lab is operated by Lawra.io for CEMI Labs, under the “augmentation, not replacement” conviction the family shares.
Comments
Loading comments...