Frequently asked questions

Each statement ends in a citation. Clicking it opens the document in the right panel with the supporting passage highlighted, and the panel header gives the source family, the document, its type and its date. Under the question, "Evidence gathered" lists every search and lookup the answer ran, with the number of steps.

Litigators on both sides and the experts they retain, pharma and device safety, regulatory and legal teams, product-liability insurers, litigation funders, M&A and licensing diligence teams, investment analysts, payers, hospital risk managers, consultancies, journalists and NGOs.

Working product. Test links on request: write to hello@kelsey.legal.

31 public sources:

• FDA adverse-event reports (FAERS)
• labels and approvals: DailyMed label versions, Drugs@FDA, the Orange Book
• FDA reviews and actions: approval reviews, safety label changes, safety communications, warning letters, advisory committee material, citizen petitions, guidance
• regulations and statutes: eCFR, the US Code, the Federal Register
• studies and trials: PubMed, ClinicalTrials.gov
• regulators abroad: EMA, MHRA, Health Canada and Swissmedic actions, and Canada Vigilance reports, kept out of the US counts
• devices: MAUDE, 510(k), PMA, recalls with FDA's class, GUDID
• company and prescription data: SEC filings, Medicare Part D, Medicaid
• courts: federal and state court decisions, state product-liability statutes
• vocabularies: RxNorm, ATC

The Sources page lists each one.

A nightly pipeline checks for updates; refresh schedules and coverage vary by source, from nightly to weekly, monthly, quarterly or yearly, and some sources are closed archives. Anything missing from the stored data is fetched live from PubMed, DailyMed and ClinicalTrials.gov. Each figure carries "Data as of" and its date.

Each date carries its type. Published is when a study or document appeared. Submitted is when a company filed something with FDA. Approved is when FDA acted. Where FDA held a document before anyone else could see it, the answer reads "Document dated [date]; public from [date]". The Methodology page has the detail.

Counts are one report per FAERS case id. Likely duplicates across case ids are found by record linkage, the method behind WHO-UMC's vigiMatch, and the signal card shows how the measures move without them. Lawyer-filed reports are flagged and kept out of the measures by default; the card gives their number, and the signal can be shown with them included.

Many of them. Each figure on a calculation card comes from a stored query run with its data snapshot and filters. For a concept such as "eye damage", the answer lists the terms that make it up. The Expert pack holds VERIFY.md, the command that opens the named snapshot read-only, reruns each eligible query with its recorded arguments and compares the numbers: counts exactly, decimals to a relative difference of 1e-9. It needs the code version named in the appendix and the data snapshot. It compares each exhibit's data, not every number in the written answer. Document, case-law and literature cards, and live lookups, are listed as not checked, because their figures do not come from a snapshot.

• The answer as PDF or Word, with the methods appendix.
• The Expert pack, a zip: exhibits.json (every card in the answer, each stamped with its data snapshot, code version and method version), the methods appendix in Markdown and PDF, and VERIFY.md.
• A draft memo or email written from the cards on screen. You edit it in the app and download it as Word on your office letterhead.

Every finding carries one of four states:

• Found
• None found; the sources searched are complete for a stated period
• Can't tell from this search, with the reason, for example "no records of this source cover this drug" or "the records searched are incomplete for this period"
• Sources disagree

An empty literature search reads "The PubMed search for [terms] through [date] returned no studies." An earliest date reads "Earliest public evidence found" only when the source was searched completely; otherwise "Earlier public evidence found ... so an earlier one may exist".

Choose a drug or a condition, optionally a side effect, and a look-back window, then press Start watching. Each check compares the result with the previous check and adds one alert listing the evidence added or removed; the first check sets the baseline. Alerts show on the Watch lists page, marked "New" with a count, together with the date the list was last checked. Alerts appear in the app; none are sent by email. A watch list never predicts or ranks.

Yes: MAUDE device reports, 510(k) clearances, PMA approvals, recalls with FDA's class, and GUDID. A question can ask, for example, whether the predicate devices a product was cleared against had the same failure.

EMA, MHRA, Health Canada and Swissmedic actions are held with their dates, so an answer can show whether a foreign regulator warned before FDA, and by how long. Canada Vigilance reports are kept out of the US counts.

No. Kelsey sets out sources, dates and their limits; it does not assess a claim or advise on one. A legal memo carries the line "Written by AI from the exhibits below. Check the numbers against them."

Kelsey is operated by Original Media BV, a company in Kalmthout, Belgium, until its US company is formed. See the legal notice.

The name is Frances Oldham Kelsey's, the FDA medical officer who refused to approve the US application for thalidomide.

No.

Write to hello@kelsey.legal with your firm or company, your role, the drug or device and the question you are working on, and the court or MDL if there is one. Leave out confidential case details.