How to do a Daubert workup in 5 seconds (instead of 5 hours)

By MedLegal AI Editorial · 9-minute read · Published April 25, 2026

Friday, 4:55 PM

The defense's expert disclosure hits your inbox four minutes before the close of business. The deposition is Tuesday morning at 9:00. The expert is a 62-year-old interventional cardiologist out of an academic hospital you do not practice near, and the disclosure is a 28-page CV with 140 publications, four prior expert reports referenced by case caption only, and a one-paragraph summary of the opinion he intends to offer.

You have, by the time the weekend ends and you account for travel, roughly 36 working hours to figure out who this person is. What he has been paid by which device manufacturers. Whether he has testified for the defense in cases like yours and what he said. Whether his publications support or quietly contradict the opinion he is about to offer. Whether he has been disciplined by any state board, named in any prior litigation, or quoted in any deposition transcript that lives somewhere a search engine can find it.

This is the work that wins or loses the Daubert motion. It is also the work that, in most plaintiff firms, gets done by a paralegal on a Saturday afternoon with eleven browser tabs open and a yellow legal pad. This post walks through what that paralegal is actually doing across nine public data sources, what kind of impeachment material lives in each, and what changes when the same nine sources get hit in parallel by software in five seconds instead of five hours by a human in series.

The 9-source manual workflow plaintiff paralegals run today

A competent Daubert workup on a defense medical expert pulls from at least nine public data sources. Each one has its own quirks, its own search syntax, and its own kind of cross-examination material. The order below is roughly the order an experienced plaintiff paralegal will hit them, because some sources feed queries to others.

1. PubMed (NCBI / NLM)

The starting point. Search the expert's name as author, restricted to publications in the last fifteen years. You are looking for two things: what has the witness actually published, and does any of it cut against the opinion he is about to give. The number-one Daubert-step-1 attack on a medical expert is "your own published work says the opposite." Pull the abstracts, flag any review articles or editorials (those carry more cross-examination weight than case reports), and note co-authors. Co-authors who are themselves frequent defense experts is a pattern worth knowing.

2. CMS Open Payments (Sunshine Act database)

Federal law requires every drug and device manufacturer to report payments to physicians. The CMS Open Payments database lets you search by NPI and pull every consulting fee, speaking honorarium, research grant, and meal the witness has received from industry, by year, by company. For a cardiologist defending a stent case, "you received $186,000 in consulting fees from the manufacturer of the device whose use you are now defending" is bias-cross territory under Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993), reliability-factor analysis — specifically the question of whether the expert's methodology is the product of independent inquiry or commissioned advocacy. Open Payments is also the single best source for the question every juror cares about: who pays this witness, and how much.

3. NIH RePORTER

The federal grants database. Search by PI name and you get every NIH-funded study the expert has been principal or co-investigator on. Two uses. First, did he take federal money to study the very procedure he is now opining about, and does the grant abstract describe limitations he is now papering over. Second, the absence of grant funding for an academic in a research-active specialty is itself a soft credibility data point. NIH RePORTER also surfaces co-investigators, which feeds back into your PubMed co-author analysis.

4. CourtListener federal dockets (Free Law Project)

The single richest impeachment source on the public internet. CourtListener's RECAP archive contains millions of federal docket entries, including expert reports filed as exhibits, deposition excerpts attached to motions, and Daubert orders that cite the expert by name. A search for the witness's name across federal dockets will surface every prior case in which his testimony was challenged, every order excluding or limiting his opinions, and frequently the prior expert report itself as a PDF exhibit. Prior excluded testimony is the highest-value impeachment material that exists. CourtListener is where it lives.

5. FSMB and individual state medical boards

The Federation of State Medical Boards aggregates licensure data, but the high-value disciplinary material lives on individual state board sites. Pull the witness's primary licensure state and any state where he holds an active license. Look for board orders, consent agreements, voluntary surrenders, and probationary terms. State boards also publish malpractice settlement reporting in some jurisdictions (Florida and Massachusetts are the two most useful). A board action is rarely admissible at trial without a fight, but it is always admissible in the Daubert qualifications attack on whether this person should be testifying at all.

6. SEC Form 4 filings (EDGAR)

Often missed. If the expert sits on a scientific advisory board for a publicly traded medical device or pharmaceutical company, his stock awards and option grants are reported on Form 4 within two business days of the transaction. Search EDGAR by name. The cross-examination is "you own $400,000 in restricted stock units in the company that makes the device you are defending today." That fact almost never appears in the expert disclosure or the CMS Open Payments database (which does not capture equity grants from publicly traded employers).

7. IRS Form 990 employer filings

Every nonprofit hospital and academic medical center files an annual Form 990 that lists the highest-compensated employees by name and total compensation. If the expert is on staff at a 501(c)(3) hospital, his Form 990 line is a public number. The use is not the dollar figure on its own. It is the comparison: an expert charging $1,200 per hour for litigation work whose institutional salary is $480,000 is making meaningful additional income from testifying. That is a frequency-of-testifying foundation question.

8. Medicare Provider Utilization and Payment Data

CMS publishes annual provider-level data on Medicare billing: how many of each procedure each NPI billed, what was paid, what was charged. For an interventional specialist, this answers "how many of the procedure at issue does this witness actually perform per year, and is that number consistent with the 'I do hundreds of these' claim he is about to make at deposition." Volume claims are routinely overstated in expert disclosures and routinely impeachable from CMS data.

9. ProPublica Dollars for Docs and Nonprofit Explorer

Layered on top of CMS Open Payments and the Form 990 filings, ProPublica provides cleaner search interfaces and longitudinal views that the underlying government databases do not. Useful primarily as a sanity check on the CMS Open Payments numbers and as a faster path to the Form 990 employer line. Free, well-maintained, and increasingly the source plaintiff paralegals reach for first even when CMS is the underlying data.

A thorough manual run across these nine sources, with notes captured into a working memo and exhibits saved into a case folder, takes a competent paralegal four to six hours. On a Friday-disclosure Tuesday-deposition timeline, that is the entire weekend.

What automation actually changes

The honest description of what software is doing here is not "AI does the Daubert workup for you." It is two distinct things stacked on top of each other.

The first layer is parallel HTTP fetches. The nine sources above all expose either a public API or a structured search endpoint. PubMed has E-utilities. CMS Open Payments has a JSON API. NIH RePORTER has a query endpoint. CourtListener has a documented REST API. FSMB and state boards are mostly screen-scrapes against stable URLs. EDGAR is a structured XBRL feed. The Medicare provider data is a quarterly downloadable dataset. ProPublica has a Nonprofit Explorer API. When the human paralegal hits these in serial, total wall-clock time is dominated by the page-load and read-and-take-notes loop. When software hits them in parallel with a known NPI and a known name, the network round-trips finish in roughly two to four seconds for the slowest source, and the rest of the data is back faster than the slowest one.

The second layer is the cross-reference. A pile of returned JSON from nine sources is not a Daubert memo. The work that turns raw data into impeachment material is matching: matching the device manufacturer in the CMS Open Payments line items against the device the expert is defending in this case; matching co-authors on his PubMed publications against named experts on the defense side; matching his prior CourtListener testimony against the opinion he is offering today and surfacing any inconsistencies; matching his Medicare procedure volumes against the volume claim in his expert disclosure. This is what a language model is genuinely good at — pattern-matching across several heterogeneous text sources and surfacing the contradictions a human reader would have caught if a human reader had time to read all of it.

What is not happening, and what we will not claim is happening, is the model "deciding" what is a Daubert vulnerability. The model produces a structured memo with citations. The attorney decides what to do with it. The five-second number is the time from clicking the button to having the memo on the screen. The decision-making time is whatever you bring to it.

The 14-section auto-generated Daubert memo

The memo Courtroom AI produces is not a chat-style summary. It is a structured document with fourteen named sections, each populated from specific sources, each citing those sources by URL. The sections, in order:

1. Witness identification — full legal name, NPI, state license numbers, primary affiliation.
2. Qualifications summary — degrees, board certifications, current academic appointment, hospital privileges.
3. Publication record — PubMed-indexed work last 15 years, with abstracts, with co-authors flagged.
4. Industry payments — CMS Open Payments by year and manufacturer, with totals and trend.
5. Equity and SAB positions — SEC Form 4 holdings and scientific advisory board roles where disclosed.
6. Federal grant funding — NIH RePORTER awards as PI or co-I, with project topics.
7. Procedure volume — CMS Medicare utilization data for the procedure at issue.
8. Disciplinary history — FSMB and primary-state-board records, board orders, consent agreements.
9. Prior testimony — CourtListener and RECAP search for prior expert reports and deposition excerpts.
10. Prior Daubert challenges — cases where this witness was challenged, with rulings and citations.
11. Methodology cross-check — the expert's published methodology vs. the methodology underlying the present opinion.
12. Bias indicators — payment-to-defense-counsel patterns, frequency-of-testifying ratio, plaintiff-vs-defense split.
13. Suggested cross-examination outline — foundation, qualification, methodology, and bias question sets.
14. Open questions — specifically what the system could not find and where a human researcher should look next.

The difference between this and a generic ChatGPT prompt is the difference between citation grounding and confident-sounding fabrication. Every factual line in the memo carries a source URL. The publication list is generated from the actual PubMed JSON return, not from the model's training data. The Open Payments numbers are pulled from the live CMS API, not summarized from a remembered news article. When the system has nothing to say about a section, it says so in section 14, rather than making something up.

That is the part to insist on with any AI tool a plaintiff firm uses. The output has to be checkable, line by line, against a primary source. If you can click the citation and land on the source document, the workflow is sound. If the model is producing prose that sounds right but does not link out to anything, you are looking at a hallucination risk that does not belong in litigation work.

Live citation verification during the deposition

The second piece worth describing is what happens at the deposition itself. When the witness, mid-testimony, says "well, the literature on this question goes back to Smith v. Jones and the consensus has been settled since then," the live system does two things in real time.

First, it pulls the actual opinion from CourtListener using the case name as a query. Second, it cross-references the holding the witness just summarized against the holding the opinion actually contains. If the witness has flipped procedural posture, mischaracterized the holding, or cited a case that says the opposite of what he claims, that contradiction surfaces on the attorney's screen within five to ten seconds of the witness finishing the sentence.

The worked example we use in training is a tissue plasminogen activator (tPA) saddle pulmonary embolism case where a defense expert cited a prior federal case for the proposition that systemic tPA was contraindicated in the patient's hemodynamic profile. The actual cited opinion held the opposite — that the defendant physician's failure to administer tPA in a comparable hemodynamic profile was not protected by the standard of care. The contradiction surfaced on the screen at minute thirty-two of the deposition. The follow-up question wrote itself: "Doctor, would you mind if I read you the holding from the case you just cited."

That is what real-time citation verification buys you. It does not buy you the cross-examination instinct to use the contradiction. It buys you the surfacing of the contradiction in time to use it. Compare to the alternative, which is realizing on Wednesday afternoon, while reading the rough transcript, that the expert misstated the holding on Tuesday morning, when the witness has already left the room.

What this does not replace

A few honest limits, because confidence in the tool depends on knowing where the tool ends.

It does not replace the attorney's judgment about which threads to pull. A 14-section memo that flags eleven potential vulnerabilities is not telling you which two will work in front of this judge in this jurisdiction with this jury pool. That is your job and it always will be.

It does not read between the lines of an expert's testimony. The system catches contradictions that exist in text. It does not catch the witness shading a non-answer, the slight pause before "I don't recall," or the body language that tells an experienced trial lawyer the witness is hiding something. Those are human signals and they require a human in the room.

It cannot subpoena prior testimony that is not already in CourtListener or RECAP. State court depositions, sealed federal exhibits, and arbitration testimony are largely invisible to public-database tooling. If you suspect a witness has testified in a state court case that is not on CourtListener, you still need to issue the subpoena and pay the court reporter for the transcript. The tool widens the public-record search dramatically. It does not replace targeted formal discovery.

And it cannot decide, ultimately, whether to file the Daubert motion. That is a strategic question about the judge, the timing, the rest of the defense case, and what gets surrendered if the motion is denied. The memo gives you the material. The decision is yours.

Try it on your next defense expert disclosure

The fastest way to evaluate whether this fits your practice is to run it once, on a case you already know cold. Take the last defense expert disclosure your firm received, run it through the workup, and compare the auto-generated memo against the manual workup your paralegal already did. The places the memo agrees with your existing notes are confirmation. The places the memo surfaces something your manual workup missed are the reason to use it on the next case. Trial signups and a 50% three-month pilot code are at medicalai.law/courtroom-ai.