The legal AI model the network trains. It is fine-tuned on knowledge licensed attorneys contributed on purpose, improved by the corrections those attorneys make on real work, and constrained by rules it does not get to argue with. It does the reasoning inside every endpoint FinePrint runs.
The definition
OpenLegalLM is seven things working together: a frontier base model, a fine-tune on legal knowledge that licensed attorneys contributed under agreement, RLLF — reinforcement learning from legal feedback — retrieval over your own company record, deterministic endpoint rules the model cannot override, evaluation suites that decide whether a change is an improvement, and release gates that make every version something you can point at. A chat model is only the first item on that list.
That distinction is the whole product. A model on its own is confident and unaccountable. Wrapped in retrieval it knows your facts. Wrapped in rules it cannot reason past a signature requirement or a filing deadline. Wrapped in evals and gates it can be improved on purpose, in versions, by people whose names are on the corrections.
A frontier lab can buy compute. It cannot buy the corrections of practicing attorneys marking up real transactions, contributed with rights and consent attached. That is the part that has to be earned, one review at a time.
// what runs when an endpoint reasons
◆Base — a frontier model, swapped when a better one ships. Rides the bases rather than racing them.layer 1
◆Fine-tune — contributed playbooks, positions and clause patterns, rights-cleared and attributed.layer 2
◆RLLF — accepts, rankings, redlines, rejections and escalation calls from licensed reviewers.layer 3
◆Retrieval — your company record, isolated to your company, per-document keys.layer 4
◆Rules — deterministic endpoint logic: required facts, approvals, signature order, escalation conditions.layer 5
◆Evals — regression, jurisdiction and escalation suites run against every candidate.layer 6
◆Gates — governed, versioned releases. One correction never changes production.layer 7
✓Two corpora, kept apart: contributed knowledge trains the model · your record is retrieved and never trained on.
Not a benchmark claim. This card describes what runs, not how well it runs. The one number we intend to publish is the flag rate — and it ships with the eval that produces it, not before.
The interesting question about a legal AI model isn’t what it can produce. It’s what stops it producing something confident and wrong. Every layer here exists because of a particular way that happens.
A frontier model does the language work: reading long documents, holding structure, drafting in register. We swap it when a better one ships. Prevents spending the network’s advantage on a capability race we would lose — the reasoning floor rises for free, and the legal layers ride on top of it.
Trained on what attorneys in the OpenLegal network contributed on purpose: playbooks, negotiating positions, clause patterns, the fallbacks that actually get accepted, and the reasons behind each. Prevents the failure mode of a general model on legal work — fluent, plausible, and subtly wrong about which position is market and which is a giveaway.
Reinforcement learning from legal feedback. Not thumbs up and thumbs down — the structured judgment of licensed reviewers: what they accepted, how they ranked alternatives, what they redlined and why, what they rejected outright, and which matters they said should never have run without a lawyer. Prevents improvement by opinion. Every signal comes from a professional who is accountable for it.
At run time, the model reads the Company Legal Graph and the Legal DataRoom — your charter, your board authority, your cap table, your agreements — isolated to your company, encrypted, with per-document keys. Prevents generic answers about a generic company, and prevents one customer’s facts ever reaching another’s matter.
Each endpoint carries logic that is code, not inference: the facts required before it may start, the approvals that make a document valid, the signature order, the filing deadlines, and the conditions that make a matter Yellow or Red. Prevents the most dangerous failure a legal AI model has — reasoning its way past a rule because the reasoning sounded good.
Regression suites built from every past correction, jurisdiction suites that check the same task across the states it has to be right in, and escalation suites that test whether the system still knows when to stop. Prevents the fix that quietly breaks something else — the reason a model can feel better and be worse.
Nothing reaches production because it looked better in a demo. A candidate has to clear the suites, and releases are versioned so a change in behavior has a name, a date and a diff. Prevents a single reviewer’s correction from changing what happens to every company — which is the difference between a model that improves and a model that drifts.
The pipeline
Generate → Review → Curate → Improve → Test → Release. Six steps between an attorney marking up a clause and that judgment changing what an endpoint does — deliberately six, because the alternative is a system that learns from whoever complained last.
01 · Generate
An endpoint produces real output on a real matter: a draft, a redline, an issue list, a level. Nothing is written for training. The corpus is a by-product of work somebody needed done, which is why it looks like practice rather than like a dataset.
02 · Review
Green output goes to the network anonymized through the Clean Room pipeline, sampled and scored. The reviewer accepts, ranks alternatives, redlines, rejects, or escalates — and names the governing reason for whichever one they chose. Reviewers are paid per review, never contingent on approving anything.
03 · Curate
Corrections are de-duplicated, checked against each other, and weighted by the reviewer’s standing in that area of practice and that jurisdiction. Disagreements between reviewers aren’t averaged away — they’re routed for a second opinion, because two attorneys differing is information.
04 · Improve
Some corrections belong in the weights. Many don’t: a missed escalation becomes a deterministic rule, a wrong position becomes a playbook edit, a jurisdiction gap becomes a retrieval fix. Choosing the right layer is most of the work, and it’s why the model isn’t asked to remember things code should enforce.
05 · Test
Regression against every correction already made, jurisdiction checks across the states the task has to be right in, and escalation tests that ask whether the system still stops where it should. A candidate that improves drafting and loses an escalation does not ship.
06 · Release
What passes ships as a version, and the endpoints move to it together. A single correction never changes production on its own — it changes a candidate, which then has to earn its way through the gate like every other change.
Step two is where the quality actually comes from, so it’s worth being precise about what a reviewer does. There are five things they can do, and each one is a different kind of signal.
| Reviewer action | What it means | What it can change |
|---|---|---|
| Accept | Correct as drafted, for these facts and this jurisdiction. The reviewer is putting their standing behind it. | A positive signal, weighted by standing. Accepts alone never ship a change — they’re how a candidate proves it didn’t regress. |
| Rank | Two or more acceptable drafts, ordered. Which position to lead with, which fallback to hold, which clause reads better to a counterparty. | Preference data — the signal that moves drafting from defensible to good, and the hardest kind to get anywhere else. |
| Redline | Wrong as drafted, and here is the language it should have been. The most expensive signal to produce and the most valuable to have. | A paired example: what was produced, what it should have been, and the reason. Usually a fine-tune signal and a playbook edit at once. |
| Reject | Not fixable by editing — wrong approach, wrong instrument, wrong law. | A negative example, plus a new case added to the regression suite so no future release can reintroduce it quietly. |
| Escalate | This should not have run without a lawyer. The output may even be fine; the level was wrong. | A proposed change to the endpoint’s deterministic escalation conditions — a rules change, not a weights change. The system learns to stop sooner. |
| The governing reason | Required with every action above: statute or regulation, the client’s own playbook, market practice, or a drafting error. | Makes each correction auditable and lets curation weight it properly. A correction without a reason is an opinion, and opinions don’t train anything here. |
The Two-Corpus Rule
Most AI companies ask you to accept a trade: give us your data, and the product gets better for everyone including you. We think that trade is unnecessary for legal work, and we built the system so we never have to ask for it.
Anonymized output goes out for review. Corrections and revenue share come back. Your record never leaves the left-hand side.
Legal AI is full of numbers that mean nothing: accuracy scores against a private test set, pass rates on an exam nobody practices with, percentages with no denominator. We would rather have one number that is hard to fake.
How often the network’s reviewers correct the model’s work: the share of sampled output that a licensed attorney accepts without change, against the share they rank down, redline, reject or escalate. It is measured on real matters, by professionals paid per review whose pay does not depend on the answer.
A falling flag rate is quality you can watch over time — and because reviewers are independent of the application side and cannot have a flag overridden, it is a number the company that publishes it cannot quietly improve.
A rate without a method is decoration. Before the number means anything you have to know what was sampled and how, which categories and jurisdictions are in scope, how a disagreement between two reviewers is resolved, and what counts as a flag rather than a preference.
So the number and the evaluation that produces it publish together. Until then this page describes mechanism and declines to describe performance — which is the honest version of a claim, and the only one worth reading.
—
Flag rate. Publishes with the eval, not before.
1
One public number, with its method attached. Not a leaderboard.
0
Benchmarks claimed, accuracy scores quoted, or exams cited. None until an eval exists.
6
Steps between one attorney’s correction and a production release.
The model works. The network judges. Everything below is a constraint we built in on purpose, and each one costs us something — which is how you can tell it’s real.
| It does not | Why the constraint exists | What happens instead |
|---|---|---|
| Give legal advice | FinePrint is a legal technology company, not a law firm. Advice is a professional relationship with a licensed person on the other end of it. | Where a matter requires legal judgment, a licensed attorney reviews it or takes it — engaged by you, with the scope shown before any work begins and the review included in your plan. |
| Hold a judgment call | A judgment call is a decision somebody has to be accountable for. A model cannot be accountable, and confidence is not accountability. | The model can say a term is outside your playbook, name the usual fallback, and price the review. The call goes to a person, with the file already built. |
| Override a reviewer’s flag | If the application side could overturn a flag, the flag rate would measure our patience instead of the model’s quality. | A flag stands. It routes to curation, and where it is an escalation it becomes a proposed change to the endpoint’s deterministic rules. |
| Override a deterministic rule | Required facts, approvals, signature order, filing deadlines and escalation conditions must not be defeasible by an argument, however good it sounds. | Those live in code outside the model. A run that cannot satisfy them stops, and the reason is written to the matter. |
| Train on your record | The Two-Corpus Rule. A model that learns from customer files is one configuration error away from a disclosure it cannot take back. | Your record is retrieved for your matters and nothing else. Contributed knowledge and reviewer judgment train the model, in a separate account. |
| Ship on a single correction | One attorney is not the profession. A system that updates on the last opinion it heard drifts, and drift is invisible until it isn’t. | Corrections are curated and weighted, changes are tested against regression, jurisdiction and escalation suites, and releases are versioned. |
| Claim a benchmark | There is no accepted public benchmark for the work these endpoints actually do, and a private one you grade yourself is marketing. | Mechanism now, the flag rate when the eval that produces it is published. Nothing in between. |
Every set of model weights has a shelf life. A better base ships, the frontier moves, and last year’s fine-tune is a liability rather than an asset. Anyone whose advantage lives in a particular set of weights is renting it.
What doesn’t melt is the corpus underneath: the contributed playbooks, the reviewed transactions, the corrections with reasons attached, the regression cases from every mistake already made, the escalation conditions learned the hard way. That material transfers to whatever base comes next, and it is larger every month than it was the month before.
So OpenLegalLM rides the frontier bases rather than racing them. When a better base arrives we take it, and the seven layers above it are the same seven layers, now sitting on a better floor. The compute was never the scarce thing.
Agents run the work. Attorneys review it. The reviews train the model. The model runs the work better — and more attorneys join a network that is worth more to belong to. The loop is the moat.
No, and we don’t intend to. OpenLegalLM sits on a frontier base model and adds the layers that make it useful and safe for legal work: the fine-tune, RLLF, retrieval over your record, deterministic rules, evals and release gates. When a better base ships, we swap it — the legal layers move across.
It rides the frontier bases rather than racing them. Racing them is expensive, and it isn’t where the difficulty in legal work lives.
Reinforcement learning from legal feedback: the same family of technique as learning from human preferences, with the preference signal restricted to licensed attorneys acting in their professional capacity, on real work, with a governing reason recorded for every action — statute, playbook, market practice, or drafting error.
The difference that matters isn’t the algorithm. It’s who is allowed to produce a signal, what they have to say when they produce it, and the fact that escalation signals are routed into deterministic rules rather than into the weights.
Into the endpoint’s rules, not the model. When a reviewer says a matter should never have run Green, curation turns that into a proposed escalation condition — a change to code that decides levels. It ships through the same gates, and afterwards the endpoint stops sooner for every company, whatever the model happens to think.
Retrieval is scoped to your company, your record is encrypted with per-document keys, and customer data and learning data live in two separated AWS accounts that never share one. Your documents are never training material for a shared model. The record is yours to export or permanently delete, any day you choose.
It arrives with provenance, consent and a permitted use, and it stays attributed. You’re paid per review, and you share in the revenue the knowledge you contributed generates, for as long as it keeps working. Your reviews train the model; the application side cannot override a flag you raise; and your pay is never contingent on approving anything.
Every correction the network has ever made becomes a regression case. A candidate release has to clear those, plus jurisdiction suites that check the same task across the states it must be right in, plus escalation suites that test whether the system still stops where it should. A candidate that drafts better and escalates worse does not ship.
With the evaluation that produces it, and not before. The eval has to state the sampling method, the categories and jurisdictions in scope, how reviewer standing is weighted, how disagreement between reviewers is resolved, and what counts as a flag. A number without that is decoration, and we would rather be the company that said so.
Which is why the network is the part we build hardest.
See the model doing the work. Six specialist agents, forty-one endpoints, and the deterministic rules that decide where an agent stops and a lawyer starts.
Inside the agentsOr look at the two things this page rests on: the licensed attorneys whose reviews train the model, and the isolation that keeps your record out of it.
