FinePrint LM. The Legal AI model.
One Legal AI model for every company on FinePrint, trained by lawyers through AnswerLoop. One for your company alone: FinePrint LM for Enterprise.
The definition
A Legal AI model, trained by lawyers.
Fine-tuned on playbooks, positions and clause patterns that legal professionals contribute. Every question a lawyer answers through AnswerLoop, every correction, redline and stop, trains the model.
It gets better with every matter. Work that needed a lawyer moves to a confirmation, then runs on its own. Red → Yellow → Green.
// what runs inside every endpoint
◆Base — the best frontier model, swapped when a better one ships.layer 1
◆Fine-tune — legal knowledge contributed by licensed attorneys.layer 2
◆AnswerLoop — learning from lawyers’ answers on real matters.layer 3
◆Retrieval — your Legal Room, isolated to your company.layer 4
◆Rules — endpoint logic the model cannot override.layer 5
◆Evals — every release tested before it ships.layer 6
◆Gates — every release versioned before it ships.layer 7
✓Two corpora, kept apart: lawyers’ answers train the model · your record is read and never trained on.
What runs, not how well it runs. The numbers we publish ship with the evaluation that produces them.
Seven layers. Each one stops a specific failure.
This is what runs inside every endpoint. Open a layer for the failure it stops.
01 · Base — the best frontier model, swapped when a better one ships
A frontier model does the language work: reading long documents, holding structure, drafting. We swap it when a better one ships. Prevents a race we do not need to run. The legal layers ride on top of it.
02 · Fine-tune — legal knowledge contributed by licensed attorneys
Playbooks, negotiating positions, clause patterns and the fallbacks that get accepted, contributed by licensed attorneys. Prevents the failure of a general model on legal work: fluent, plausible, and wrong about which position is market.
03 · AnswerLoop — learning from lawyers’ answers
Not thumbs up and thumbs down. The answers of licensed lawyers to questions the model had not seen: what they agreed with, modified, reworked, or said should never have run without a lawyer. Prevents improvement by opinion.
04 · Retrieval — your Legal Room, isolated to your company
At run time the model reads your Legal Room and Legal Room. Encrypted, company-scoped, per-document keys. Prevents generic answers about a generic company, and one customer’s facts reaching another’s matter.
05 · Rules — endpoint logic the model cannot override
Each endpoint carries logic that is code, not inference: required facts, the approvals that make a document valid, signature order, filing deadlines, and the conditions that make a matter Yellow or Red. Prevents the model reasoning its way past a rule.
06 · Evals — every release tested before it ships
Regression suites built from every past answer. Jurisdiction suites that check the same task across states. Escalation suites that test whether the system still knows when to stop. Prevents the fix that breaks something else.
07 · Gates — every release versioned before it ships
A candidate has to clear the suites. Releases are versioned, so a change in behavior has a name, a date and a diff. Prevents one correction from changing what happens to every company.
How it learns
How an answer becomes a release.
AnswerLoop, step by step. Six steps between a lawyer answering a question and that answer changing what an endpoint does. Every answer moves work one lane to the right.
- Generate. An endpoint produces real output on a real matter. Nothing is written for training. The data is a by-product of work somebody needed done.
- Answer. A licensed attorney answers the Work Request: agree, modify or rework, and names the governing reason each time. An answer without a reason does not train anything.
- Curate. Answers are de-duplicated and weighted by the reviewer’s standing in that practice area and jurisdiction. Disagreement between reviewers goes to a second lawyer.
- Improve. The change is made where it belongs: a missed escalation becomes a rule, a wrong position a playbook edit, a drafting failure a fine-tune signal.
- Test. Candidates face regression against every past answer, jurisdiction checks, and escalation tests. A candidate that drafts better and escalates worse does not ship.
- Release. What passes ships as a version. The endpoints move to it together. The released model runs the next endpoint, and the loop turns again.
Your Legal Room
FinePrint LM learns from lawyers’ answers, never from your Legal Room.
The Two-Corpus Rule: FinePrint LM learns through AnswerLoop from legal professionals, never from your Legal Room.
Corpus one — what trains the model
- Contributed knowledge. Playbooks, positions and clause patterns, with rights, consent and a permitted use attached to each source.
- Lawyers’ answers. The agreements, modifications and reworks made on anonymized output.
- Public law. Statutes, regulations, forms and filing requirements.
Corpus two — what is only ever read
- Your Legal Room. Read for your matters. Encrypted and company-isolated with per-document keys.
- Your documents. Never training material for a shared model. Not aggregated, not sampled, not “de-identified and reused.”
- Two separate accounts. Customer data and learning data live in two AWS accounts. The rule is enforced by architecture, not by policy.
Anonymized output goes out for review. Answers and revenue share come back. Your record never leaves the left-hand side.
FinePrint LM for Enterprise
One model for everyone. One for your company.
FinePrint LM for Enterprise is your company’s own secure, private Legal AI model, trained by your legal knowledge, used by FinePrint AI counsels and never shared with anyone outside.
| FinePrint LM | FinePrint LM for Enterprise | |
|---|---|---|
| Serves | Every company on FinePrint. | One enterprise. |
| Built from | Legal knowledge from professionals, fine-tuned into the Legal AI model. | FinePrint LM, plus your contracts, playbook and Legal Room. |
| Trained by | A network of approved legal professionals, through AnswerLoop. | Your own lawyers, through AnswerLoop. |
| Learns from | Every reviewed matter across FinePrint. | Only your company’s matters. |
| Lives | Inside FinePrint. | Under your keys, on your terms. Exportable. |
Nothing FinePrint LM for Enterprise learns enters FinePrint LM. Nothing a company puts in leaves.
Two numbers. The flag rate and the novelty rate.
Both measured on real matters. Both publish together with the evaluation that produces them.
The flag rate
How often lawyers correct the model’s work, measured on real matters by reviewers whose pay does not depend on the answer. A flag cannot be overridden, so the company that publishes the number cannot improve it by hand.
The novelty rate
What share of matters reach a lawyer because the model had not seen the question. AnswerLoop working is this number falling, endpoint by endpoint. It publishes with its sampling method, the categories and jurisdictions in scope, and how disagreement is resolved.
—
Flag rate. Publishes with the eval.
—
Novelty rate. Publishes with the eval.
0
Private benchmarks or exam scores. We measure on real matters.
6
Steps between one attorney’s answer and a release.
What FinePrint LM does not do.
Each constraint costs us something. That is how you can tell it is real.
| It does not | Why | What happens instead |
|---|---|---|
| Give legal advice | FinePrint is a legal technology company, not a law firm. Advice is a relationship with a licensed person on the other end. | Where judgment is required, a licensed attorney reviews the matter or takes it. Engaged by you. Included in your plan. |
| Hold a judgment call | A judgment call needs somebody accountable for it. Confidence is not accountability. | The model names the issue and the usual fallback. The call goes to a lawyer, with the file already built. |
| Override a reviewer’s flag | If the product could overturn a flag, the flag rate would measure our patience, not the model’s quality. | A flag stands until a licensed attorney clears it: the reviewer, or a second attorney under the escalation rule. |
| Override a rule | Deadlines, approvals and escalation conditions cannot be argued away, however good the argument 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 | A model that learns from customer files is one configuration error from a disclosure it cannot take back. | Your record is read for your matters and nothing else. Training happens in a separate account, on contributed material. |
| Ship on a single correction | One attorney is not the profession. A system that updates on the last opinion it heard drifts. | Answers are curated and weighted, tested against the suites, and shipped as versioned releases. |
| Claim a benchmark | There is no accepted public benchmark for this work. A private one you grade yourself is marketing. | Mechanism now. The flag rate and the novelty rate when their eval is published. |
Questions, mostly from lawyers and engineers.
Is FinePrint LM its own foundation model?
No. FinePrint LM sits on the best frontier model and adds the layers that make it safe for legal work: the fine-tune, AnswerLoop, retrieval over your Legal Room, rules, evals and release gates. When a better base ships, we swap it. The legal layers move across.
What exactly is AnswerLoop, in machine-learning terms?
Reinforcement learning from lawyers’ answers. The same family of technique as learning from human preferences, with the signal restricted to licensed attorneys answering real questions on real work, and a governing reason recorded for every answer. Escalation signals go into rules, not into the weights, so the system learns to stop sooner.
What is FinePrint LM for Enterprise?
The private edition. Built on FinePrint LM, trained by your legal knowledge and your lawyers’ answers, used by your FinePrint AI counsels, and never shared with anyone outside. FinePrint for enterprise
Can my company’s data leak into another customer’s answer?
No. Retrieval is scoped to your company. Your record is encrypted with per-document keys. Customer data and learning data live in two separate AWS accounts. Your documents are never training material for a shared model. The record is yours to export in full or delete, any day you choose.
How do I know a new release didn’t make something worse?
Every answer becomes a regression case. A candidate release has to clear those, plus jurisdiction suites and escalation suites that test whether the system still stops where it should. A candidate that drafts better and escalates worse does not ship.
When will you publish the flag rate and the novelty rate?
With the evaluation that produces it. The eval states the sampling method, the categories and jurisdictions in scope, how reviewer standing is weighted, how disagreement is resolved, and the definition of a flag.
Every review moves work one lane to the right.
Every question is answered once by a lawyer and forever by the model. Now see it doing the work.