Skip to content
Lexyno
Back to Academy
Legal TechMay 12, 2026Updated Oct 14, 20268 min readJun LeeJun Lee

What Is Fact-Lock™? How Lexyno Reduces AI Errors in Demand Letters

What Is Fact-Lock™? How Lexyno Reduces AI Errors in Demand Letters

AI hallucination is a real, measured problem in legal documents — even in tools that promise "hallucination-free" output. Here is how Fact-Lock™ is engineered to reduce it, and why we will never claim zero.

Jun Lee
Jun Lee
Co-Founder & CTO

There's a comforting story the legal industry tells itself about AI. Yes, ChatGPT makes things up. Everyone knows that.

But the serious tools are supposed to be different: the ones built for lawyers and grounded in real databases. One major provider went so far as to market "100% hallucination-free linked legal citations."

Stanford researchers tested that promise. If you put AI anywhere near a legal document, the result is required reading. It's also why Lexyno will never make the claim you might expect from us.

Yes, and it's been measured. In a preregistered Stanford study, Lexis+ AI gave incorrect or misgrounded answers more than 17% of the time. Westlaw's AI-Assisted Research hallucinated around 33% of the time, and a bare GPT-4 baseline was wrong about 43% of the time.

The study came from Stanford's RegLab and Institute for Human-Centered AI. Researchers ran over 200 legal queries through the leading purpose-built tools and scored the answers by hand.

Two things should matter to you here. First, the purpose-built tools did make fewer errors than a raw chatbot. Retrieval-grounded design works.

Second, none of them came close to zero. In the researchers' words, the providers who marketed "hallucination-free" output were overstating their claims.

The legal industry already ran the experiment on "zero hallucination" marketing. Independent research found the claim overstated. We're not going to repeat that mistake.

Why Is a Wrong Number So Damaging in a Demand Letter?

A demand letter is a financial argument, so a wrong figure doesn't stay contained. If the specials total has a transposed digit, or a treatment date is invented, the adjuster doesn't just discount that line. They discount the whole letter's credibility, and your negotiating position goes with it.

In legal research, a hallucinated citation is dangerous because a lawyer might rely on a case that doesn't say what the tool claims. In a demand letter, the failure looks different. It's a wrong figure, and it does just as much damage.

The letter tells an adjuster what happened, what it cost, and what the claim is worth. A bill pinned on the wrong provider undercuts all three. Here, fluent, confident, wrong text is worse than no text at all.

So "the AI writes a nice draft" isn't good enough. The standard has to be this: can you check every fact against the source, quickly?

What Does Fact-Lock™ Actually Do?

Fact-Lock™ is the extraction and validation pipeline inside Lexyno's AI demand letter software. It structures the facts before any prose is written. It extracts and organizes the facts from the medical file first, then writes a letter constrained to those facts. A separate validation stage checks the draft against the records and flags anything it can't trace.

That design decision is simple to state and hard to engineer. Most AI tools do the opposite. They generate fluent text and hope the facts came along for the ride.

In practice, the demand passes through several stages instead of one generation step:

1. Draft
Structured facts from the records become a constrained first draft
→
2. Coherence
The narrative is checked for internal logical consistency
→
3. Validation
Each fact is checked back against the source record
→
4. POV
Attorney strategy and perspective are layered onto verified facts
The Fact-Lock™ pipeline separates fact-grounding from prose, then validates before strategy is applied.

What matters is that validation is a separate stage from drafting. In a single-pass tool, the model that invents a number is the same model asked to trust it.

Fact-Lock gives verification its own step and its own job: confirm that what the draft says is actually in the uploaded records. Facts that can't be traced get flagged, not smoothed over.

A demand letter draws on one case's records, not the whole body of case law. The AI isn't asked what the law says; it's asked what these records contain and how to organize them. That narrower job makes honest accuracy achievable, because every fact can be checked against the file.

The Stanford study measured something harder. The tools answered open-ended questions about the entire body of case law. They had to recall the law from a vast corpus, which is where misgrounding thrives.

Open-ended legal researchDemand letter (Fact-Lock)
Source of truthEntire body of case lawOne case's uploaded records
AI taskRecall and interpret the lawOrganize and cite present facts
Failure modeMisgrounded or invented citationsMisread or transposed figures
VerifiabilityHard — must check against all lawFast — check against the file

When the source of truth is the file in front of you, traceability stops being a marketing word. It becomes a workflow: your paralegal confirms any figure by opening the page it came from.

The Honest Standard We Hold Ourselves To

We won't tell you Fact-Lock never makes a mistake. No current AI system can say that, and the providers who said otherwise got caught.

Here's what we will tell you: every material fact in the letter should trace to its source. The structure should surface anything that doesn't, and a person should check the letter before it goes out.

What Fact-Lock is built to doWhat we don't promise
Tie every material figure and date to the page it came fromA letter with no mistakes
Structure the facts before any prose is writtenThat the model never misreads a page
Flag facts the validation stage can't traceA letter you can send without reading it
Give your team a draft it can check and controlA replacement for your paralegal or attorney

The right benchmark isn't a perfect, unsupervised AI. It's beating the manual paralegal draft: faster, with the facts grounded and checkable, while your team keeps the final word.

That's a promise we can keep. It's also one you can check.

Try Lexyno FreeGenerate a mock demand letter from your own records.

Frequently asked questions

Yes, and it is measured. A Stanford RegLab study found that purpose-built legal research tools from LexisNexis and Thomson Reuters still produced incorrect or misgrounded answers between 17% and 33% of the time, while a general model like GPT-4 was wrong about 43% of the time. Retrieval-augmented systems reduce errors substantially, but no current tool reaches zero.

Fact-Lock™ is Lexyno's architecture for grounding a demand letter in its source records. It structures and verifies the facts from the medical file before any prose is written, then validates the draft against those facts in a separate stage. The goal is that every figure and date in the output traces back to the source, making errors fast to catch rather than hidden in fluent text.

AI can produce an accurate first draft when it is constrained to the source documents and its output is validated and human-reviewed. It cannot be trusted to write a demand letter unsupervised, because fluent text can still contain a wrong number. The realistic model is AI for assembly with structured grounding, and a paralegal for final verification.

Because that claim is not honest, and the legal industry has already been burned by it — a major provider marketed "100% hallucination-free" output that independent research showed was overstated. Lexyno claims something verifiable instead: every fact in the letter is traceable to its source, so your paralegal can confirm it quickly. Verifiability is a stronger promise than an unprovable zero.

A demand letter is grounded in a finite, uploaded set of documents — the medical records and bills for one case — rather than the entire body of case law. That makes the accuracy problem more tractable: the AI is not recalling the law from memory, it is organizing and citing facts that are physically present in the file. Fact-Lock is built around that constraint.

Related posts

What's next?

Ready to see the difference in your next case?

Put these strategies to work. Try a mock demand generated by Lexyno today.

Request Evaluation Access