Start free trial →
Reproducible · Public depositions

Live case studies — MedLegal AI vs. real public depositions

Most legal AI vendors won't publish reproducible methodology. We did — twice. Both case studies below were run end-to-end against full public depositions on YouTube. Same production prompt. Same scoring code. Anyone with the transcript and a Claude API key can re-run identically.

2
Full depositions analyzed
22/22
Tolson signals fired
92/100
Crowden bait-switch flag
0
Fabricated citations
Case Study #1 · May 13

Tolson v. St. Agnes Healthcare — Maryland med-mal, missed knee dislocation

A 2-hour public deposition of a defense orthopedic expert, published by Miller & Zois as legal-education content. The analyzer caught the Daubert haymaker, the internal inconsistency on nursing SOC, the breach element admission, and the PubMed-grounded literature hit on popliteal artery injury.

Jurisdiction: Maryland Length: 2 hrs Theory: Missed dx
22/22
Signals fired
90/100
FRE 702 score
0
False positives
Read the Tolson case study → Case Study #2 · May 13

Crowden v. Emory — Georgia TAVR ghost-surgery, informed-consent breach

A 3.6-hour deposition of a cardiothoracic surgeon whose name was on the consent form for a TAVR procedure he didn't actually perform. The analyzer flagged the bait-and-switch admission at 92/100 vulnerability, surfaced cross-exam questions, and identified an FRE 702 methodology gap on HIT panel scoring.

Jurisdiction: Georgia Length: 3.6 hrs Theory: Ghost surgery / consent
92/100
Bait-switch flag
78/100
FRE 702 (HIT)
0
Fabricated cites
Read the Crowden case study → Cohort · May 15

7 Public YouTube Depositions, Dual-Pass Tested

A 7-case cohort drawn from public YouTube medical-malpractice depositions. Two test passes against each: closed-caption text fed to the analyzer directly, and live audio captured via Mac BlackHole loopback through Deepgram. Real numbers. Plus the silent production bug this test caught + fixed.

Cases: 7 (peds neuro, ER ×2, vascular surgery, triage nurse, ER nurse, PA) Modes: CC text + live audio Bug caught: 1 (markdown-fence parser)
134
Cross-exam Q's
14
Admissions
8
PubMed cites
Read the YouTube batch case study → Methodology · May 14

How to Audit a Legal AI Tool — Reproducible Methodology

A 5-question framework for evaluating any legal AI vendor's claims. Built from running our analyzer end-to-end against the Tolson + Crowden depositions. Apply the same process to evaluate any vendor (including us). Includes the reproducibility checklist + the 5 marketing anti-patterns we flag.

Type: Meta-essay Framework: 5 questions Anti-patterns: 5 flagged
138
Chunks scanned
22
Deep-analyzed
0
Fabricated cites
Read the methodology →

How we ran these

Same methodology both times. Same production analyzer prompt (the one that runs on our paying users' live depositions). Same model (Claude Sonnet 4.5). Nothing tuned per case.

  1. Pull the public transcript with youtube-transcript-api (Python).
  2. Chunk into ~90–180 second windows with a sliding overlap.
  3. Score each chunk against the signal-pattern regex set — memory failures, breach admissions, evasion, contradictions, qualifications gaps — to find the high-density windows.
  4. Run the production Courtroom AI analyzer prompt against the top-scoring 10–12 chunks.
  5. Output: structured JSON per chunk with vulnerabilityScore, crossExam.topQuestions, priorTestimony.inconsistencies, evasion patterns, FRE 702 vulnerability score, element advancement, and (for Crowden-type cases) informedConsent.ghostSurgeryRisk.

Across 22 deep-analyzed chunks + 116 lower-signal scan chunks, the analyzer produced 0 fabricated case citations, 0 fabricated paper titles, 0 fabricated witness quotes. Every signal that fired was traceable to a specific YouTube timestamp.

Run this against your own deposition

Upload a depo transcript or paste a YouTube URL. The same analyzer that ran on Tolson and Crowden runs on your case. Three free analyses on the trial — no credit card required.

Start free trial → See Courtroom AI →