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

There is a comforting story the legal industry tells itself about AI. Yes, ChatGPT makes things up — everyone knows that. But the serious tools, the ones built specifically for lawyers and grounded in real databases, those are different. One major provider went so far as to market "100% hallucination-free linked legal citations."

Stanford researchers tested that promise. The result should be required reading for anyone putting AI near a legal document — and it is the reason Lexyno will never make the claim you might expect us to make.

The Problem Is Real, and It Has Numbers

In a preregistered study from Stanford's RegLab and Institute for Human-Centered AI, researchers ran over 200 legal queries through the leading purpose-built tools and hand-scored the results. The findings were blunt: Lexis+ AI produced 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.

Two things matter here. First, the purpose-built legal tools genuinely did reduce errors compared to a raw chatbot — retrieval-grounded architecture works. Second, none of them came close to zero, and the providers who marketed "hallucination-free" output were, in the researchers' words, overstating their claims.

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

Why a Wrong Number Is Catastrophic in a Demand Letter

In legal research, a hallucinated citation is dangerous because a lawyer might rely on a case that does not say what the tool claims. In a demand letter, the failure mode is different but just as damaging: a wrong figure.

A demand letter is, at its core, a financial argument. It tells an insurance adjuster what happened, what it cost, and what the claim is worth. If the special damages total is off by a transposed digit, if a treatment date is invented, if a provider bill is misattributed — the adjuster does not just discount that line. They discount the entire letter's credibility, and with it your negotiating position. In this context, fluent, confident, wrong text is worse than no text at all.

This is exactly why "the AI writes a nice draft" is not a sufficient standard. The standard has to be: can every fact be verified against the source, quickly?

What Fact-Lock™ Actually Does

The core design decision behind Fact-Lock is simple to state and hard to engineer: structure the facts before writing the prose. Most AI tools do the opposite — they generate fluent text and hope the facts came along for the ride. Fact-Lock inverts that. It extracts and organizes the verifiable facts from the medical file first, and only then writes a letter constrained to those facts.

Concretely, the demand letter passes through a multi-stage pipeline rather than a single 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 reason this 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 treats verification as its own step with its own job: confirm that what the draft asserts is actually present in the uploaded records. Facts that cannot be traced get flagged, not smoothed over.

Why the Demand Letter Problem Is More Tractable Than Legal Research

Here is the part that makes honest accuracy achievable rather than aspirational. The Stanford study measured tools answering open-ended questions about the entire body of case law — the AI had to recall law from a vast corpus, which is where misgrounding thrives.

A demand letter is a fundamentally narrower problem. The relevant facts are not scattered across millions of cases; they are sitting in a finite, uploaded set of documents for one client. The AI is not being asked "what does the law say?" It is being asked "what do these specific records contain, and how do I organize them?" That constraint is everything.

	Open-ended legal research	Demand letter (Fact-Lock)
Source of truth	Entire body of case law	One case's uploaded records
AI task	Recall and interpret the law	Organize and cite present facts
Failure mode	Misgrounded or invented citations	Misread or transposed figures
Verifiability	Hard — must check against all law	Fast — check against the file

When the source of truth is the file in front of you, traceability stops being a marketing word and becomes a workflow: a paralegal can confirm any figure in the letter by looking at the page it came from.

The Honest Standard We Hold Ourselves To

We will not tell you Fact-Lock produces zero errors, because no current AI system does, and the providers who said otherwise were caught. What we will tell you is the standard Fact-Lock is engineered around: every material fact in the letter should be traceable to its source, the structure should surface anything that is not, and a human should verify before it goes out.

The right benchmark is not a perfect, unsupervised AI. It is outperforming the manual paralegal draft — faster, with the facts grounded and checkable — while keeping your team in control of the final word. That is a promise we can actually keep, and one you can actually verify.