Validation & error rates

A known error rate is a Daubert factor.
Almost no AI vendor can state one.

Federal Rule of Evidence 702 and Daubert list the known or potential rate of error as an explicit reliability factor. If software touched your methodology, opposing counsel can fairly ask what its error rate is. This page is our answer — every figure with the method behind it, the date it was measured, and an honest statement of what it does not cover.

Version 1.0 · last updated 2026-08-04 · we re-measure and re-publish this page when a model changes

What we have measured

Each card states the figure, the exact method, the date, and the limits. A number without a method is marketing, so we give you both — including where the method is weaker than we would like.

0.094
Measured

Deposition Trainer speech-to-text — word error rate 0.094 (9.4%)

How accurately the trainer's speech pipeline turns spoken deposition language, including medical and orthopedic vocabulary, into text.

Measured
2026-08-04, on the production deployment.
Models
Speech-to-text Deepgram nova-3-medical; speech synthesis Deepgram aura-2-orpheus-en.
Method
A fixed 32-word reference sentence written in deposition register with orthopedic and neurovascular vocabulary ("dislocated knee", "neurovascular assessment", "ACL and PCL reconstruction") is synthesized to 16 kHz mono linear16 audio, streamed frame-by-frame into the live speech-to-text socket with the exact production parameters, and the returned transcript is scored against the reference by word-level Levenshtein edit distance ÷ reference word length, after lower-casing and stripping punctuation. 0.094 corresponds to 3 word errors across 32 words.
Reproduce it
The measurement is code, not a slide: GET /api/diag-roundtrip on the Deposition Trainer backend, implemented in deposition-trainer/backend/src/server.js. The reference sentence, the streaming parameters, and the scoring function are all in that source.
What this figure does not cover. This is a synthesized-speech round trip, not a benchmark against real human courtroom audio, and it is a single utterance, not a corpus. It is self-measured, not independently audited. Synthesized speech is cleaner than a real deposition room — no crosstalk, no room noise, no accents, no speakerphone. Treat 0.094 as the floor of what the pipeline does under good conditions, not as an expected courtroom rate. The real-audio measurement is the next card, and it is not finished.
Pending
not yet published
Pending verification

Word error rate on real human deposition audio

The measurement that would actually answer "what is your error rate in a deposition?" We are not publishing a number for it, because we have not finished measuring it.

What exists
A real-audio diagnostic (GET /api/diag-real-audio) streams a 15-second chunk of the publicly published Tolson v. St. Agnes Healthcare deposition — a defense orthopedic surgeon, released as legal-education content, no protected health information — through the same production speech-to-text path and returns the transcript.
What is missing
A human-verified reference transcript of that audio to score against, and enough minutes of audio for the number to mean anything. Without a verified reference there is no error rate, only a transcript that looks good.
Commitment
When we have a human-verified reference over a meaningful sample, the word error rate goes on this page with its method and its limits — whatever the number is.
Why we are showing you an empty box. The entire point of this page is that a figure you cannot trace to a method is worthless. Publishing a real-courtroom error rate we have not actually measured would be exactly the behavior this page exists to argue against — and it is the kind of thing that gets taken apart on cross.
0
Measured & published

Fabricated citations — 0 across 138 chunks of two public depositions

The Mata v. Avianca failure mode: an AI tool inventing a case citation, a paper title, or a quote that does not exist.

Published
2026-05-14 — full write-up at How to audit any legal AI tool.
Corpus
Two public depositions analyzed end to end: Tolson v. St. Agnes (Maryland medical malpractice, ~2 hours) and Crowden v. Emory (Georgia TAVR, ~3.6 hours). 138 chunks evaluated, 22 of them deep-analyzed.
Result
Zero fabricated case citations, zero fabricated paper titles, zero fabricated witness quotes.
Reproducible
The inputs were public: the depositions' public transcripts, a published model API, the analyzer prompt, and the chunk-scoring code. The report states the inputs and the scoring rules so a third party — including opposing counsel — can re-run it and check us.
Why it holds
Architecture, not luck. Case law is only cited by calling out to CourtListener and reading the actual opinion; medical literature is only cited with a live PubMed lookup and a real PMID; prior testimony is only cited from documents the user uploaded. The tool has no path to invent an authority it did not retrieve.
What this figure does not cover. It is a self-run evaluation on two depositions, not an independent audit and not a large benchmark. Two matters is a small sample. The value is that it is reproducible from public inputs — you do not have to believe us, you can re-run it. Read the method before you rely on the number.
35/35
Measured

Citation-scrubber test suite — 35 of 35 passing

A post-processing layer that reads model output before you ever see it and flags any citation pattern it cannot tie to a source.

Re-run
2026-08-04node --test shared/hallucination-scrubber.test.js, 35 tests, 35 pass, 0 fail.
What it catches
Unverified case names, reporter citations, Bates numbers with no matching document, and malformed statutory cites. Anything it cannot verify is wrapped in a visible [VERIFY CITE] marker rather than being silently emitted as fact. The known-authority allowlist is deliberately tight: Daubert, Kumho Tire, Joiner, Frye, and the Federal Rules.
Coverage
The anti-hallucination instruction block is present in 28 server-side source files across the tool backends (verified 2026-08-04 by sentinel grep for ANTI-HALLUCINATION RULES).
What this figure does not cover. A passing test suite proves the scrubber behaves as specified on the cases we thought to write. It is not a measurement of how often a model would have hallucinated without it, and it cannot catch a fabrication that does not match a citation pattern. It is a safety net, not a guarantee.
Every
asserted
fact
Architecture — independently checkable

Page-level (Bates) citation on every asserted fact

Not a percentage — a design constraint. Every factual assertion in a MedLegal AI output carries the source file and the page it came from.

What it means
Output is emitted as [filename.pdf p14]-style page cites, PubMed PMIDs, and pinpoint cites into uploaded CVs and prior depositions. A shared verifier component scans the rendered output, counts the citations, and shows a running "Verified: X/Y citations linked to source" banner. Click any citation and the source page opens.
Why it matters on cross
It moves the question from "do you trust the software?" to "open page 14 and look." Opposing counsel can check any assertion against the record without taking the vendor's word for anything — and so can you, before the fact enters your report. That is the verification step the disclosure paragraph asks you to certify.
Daubert factor
Testability. An assertion tied to a page is falsifiable by opening the page. An assertion with no citation is not testable at all.
The limit that matters most. A citation proves the page was retrieved. It does not prove the page was read correctly, and it does not make an inference from that page correct. You still have to open every page you cite. The architecture makes verification possible; it does not perform it for you.

How to use this in a Daubert challenge

Whether you are defending your own methodology or attacking an opponent's, here is which reliability factor each item on this page speaks to — and where we do not have an answer.

Daubert / FRE 702 factorWhat on this page speaks to itHonest status
Whether the technique can be tested Page-level Bates citation on every asserted fact; the fabricated-citation evaluation re-runnable from public inputs; error-rate diagnostics that are live HTTP endpoints rather than claims in a deck. Strong. Every claim here is checkable by someone other than us.
Known or potential rate of error Speech-to-text word error rate 0.094 under synthesized-speech conditions (2026-08-04, nova-3-medical); 0 fabricated citations across 138 chunks of two public depositions. Partial. We can state a measured rate — which most vendors cannot — but the courtroom-audio word error rate is pending, and the citation figure is a two-matter sample. Say so plainly if you are asked.
Peer review and publication The audit methodology is published openly and invites third parties to re-run it against us. Weak, and we will not dress it up. Published is not peer-reviewed. There is no peer-reviewed literature on this software, and you should not suggest otherwise on the record.
Existence of standards controlling operation Fixed production parameters for the speech pipeline; a tight known-authority allowlist in the citation scrubber; a scrubber test suite that runs on every change; documented operating rules for experts on the disclosure page. Internal standards only. These are our controls, consistently applied and documented — not an external standard body's.
General acceptance in the relevant community The underlying speech and language models are in broad commercial use; page-cited, source-linked output is the direction the legal-AI field is moving after Mata v. Avianca. Do not overclaim this one. AI-assisted expert workflows are new. There is no general acceptance to assert. The defensible position is the one in the disclosure paragraph: the software organized and cited the record, and you formed every opinion — so the methodology on trial is yours, which does have general acceptance.

The strategic point: the reliability inquiry should be about your methodology, not the software's. Keeping the software on the "organize and cite" side of the line — and disclosing it — is what keeps it there.

What we deliberately do not claim

A credibility page that only lists strengths is an advertisement. These are the claims we could make but will not, because we have not measured them.

No "99% accurate." Accurate at what, on which dataset, against which reference? We publish figures with a stated method or we publish nothing.
No claim of independent or third-party audit. Every figure on this page is self-measured. We designed them to be re-runnable by someone else, which is not the same thing as having been re-run by someone else.
No courtroom-audio error rate — yet. See the pending card above. When it exists, it goes here whatever it says.
No claim that the software forms, checks, or validates opinions. It organizes and cites the record. Opinions are the expert's, full stop.
No clinical, diagnostic, or regulatory validation. These are litigation-support tools. Nothing here is a medical device claim or a clearance of any kind.
No peer review. Publishing our own methodology is not peer review, and we will not let it be characterized as such.

Questions

Why does an AI tool's error rate matter under Daubert?

Daubert v. Merrell Dow and Federal Rule of Evidence 702 list the known or potential rate of error as an explicit factor a court weighs when assessing whether an expert's methodology is reliable. If software touched the methodology, opposing counsel can fairly ask what its error rate is. Most AI vendors have never measured one, so the expert has nothing to answer with.

What is MedLegal AI's measured speech-to-text error rate?

The Deposition Trainer's speech pipeline measured a word error rate of 0.094 (9.4%) on 2026-08-04 using Deepgram nova-3-medical. That figure comes from a synthesized-speech round trip: a fixed 32-word reference sentence containing orthopedic and deposition vocabulary is spoken by a text-to-speech voice, transcribed, and scored by word-level edit distance. It is a self-measured internal diagnostic on synthesized audio, not a benchmark against real human courtroom audio. A word error rate against human deposition audio with a human-verified reference transcript is pending and will be published here when it is measured.

What is page-level Bates citation and why does it matter for a Daubert challenge?

Every asserted fact in a MedLegal AI output carries the source file and the page it came from, rendered as a clickable citation that opens the source page. That makes the tool's output independently checkable: opposing counsel does not have to take the software's word for anything, and the expert can verify each fact against the record before it enters a report. It speaks to the testability factor, because any assertion can be falsified by opening the cited page.

Has MedLegal AI been tested for fabricated citations?

Yes, and the test is public and reproducible. The published methodology report evaluated 138 chunks across two public depositions — Tolson v. St. Agnes and Crowden v. Emory — using only public inputs, and found zero fabricated case citations, zero fabricated paper titles, and zero fabricated witness quotes. The inputs, prompt, and scoring code are described so a third party can re-run it. Read the methodology →

Does MedLegal AI claim its tools are validated for clinical or diagnostic use?

No. These are litigation-support tools. Nothing on this page is a claim of clinical validation, regulatory clearance, or diagnostic accuracy, and nothing here is legal advice. The figures describe internal measurements of software behavior, with their limits stated.

How often is this page updated?

Whenever a measurement changes and whenever an underlying model changes. Speech and language models are versioned and get replaced; a figure measured on one model does not carry over to the next. Each card is dated so you can tell how fresh it is, and the version history below records what changed.

Version history

v1.0 · 2026-08-04First publication. Speech-to-text round-trip word error rate 0.094 (nova-3-medical, synthesized speech). Fabricated-citation evaluation, 0 across 138 chunks, carried forward from the 2026-05-14 methodology report. Citation-scrubber suite re-run, 35/35. Real-human-audio word error rate marked pending.
⚠ Not legal advice. This page describes internal software measurements and how they map to the reliability factors courts weigh. It is not legal advice, it does not create an attorney–client relationship, and it is not a prediction of how any court will rule on any challenge. Admissibility is decided matter by matter on the record before the court. Discuss any of this with retaining counsel before it goes into testimony or a filing.