SOURCEBROOK SAMPLE - DATA DICTIONARY
September 2026 · kevin@sourcebrook.com · sourcebrook.com

This sample has two files taken from the full dataset. The full dataset also holds FDA inspection
records (66,781 inspections at 26,911 facilities since October 2008, with Form 483 citations and
warning letters), EU/UK GMP non-compliance statements, and upcoming FDA decision dates.


1) sourcebrook_sample.csv - inspection problems disclosed by companies in SEC filings
   42 rows from 2023-2026, 29 companies. The full table covers 2001 to date (1,677 statements
   from 8,055 filings).

   record_id        stable ID of the statement (same across updates)
   accepted_at_et   when SEC accepted the filing, US Eastern time, from the EDGAR filing index
   after_close      1 = accepted at or after 16:00 ET, 0 = before
   filed            EDGAR filing date
   ticker           filer's ticker
   company          filer's name
   form             8-K, 10-Q, 10-K or 6-K (or the form an exhibit belongs to)
   event_type       FORM_483 (inspection observations), OAI ("Official Action Indicated"
                    classification), WARNING_LETTER, CRL_FACILITY (an FDA complete response letter
                    that cites a facility or its inspection), INSPECTED (inspection deficiencies
                    reported without a named outcome)
   event_month      latest month the passage names (YYYY-MM), never after the filing
   sentence         the sentence reporting the outcome, quoted from the filing
   source_url       the document the sentence is in
   plant_operator   filer = the filer's own plant; contract_manufacturer; partner_or_other_company;
                    unclear
   review_stage     application_under_review = tied to a drug awaiting an FDA decision;
                    marketed_products; unclear
   fda_fei          FDA facility ID(s) the passage matches, when it pins 1 to 3 plants
   fda_plant        FDA's legal name for that plant
   plant_match      name = the plant's legal name is in the passage; city = the city plus the
                    firm's name are


2) sourcebrook_plant_links_sample.csv - which plant makes which company's drugs
   25 rows, 20 companies, each pinned to a single FDA plant. The full table covers 655 companies
   (210 linked to at least one FDA plant today).

   ticker                     company
   drugs                      products the link applies to (as named in the source)
   role                       what the plant does (drug substance, drug product, fill-finish, ...)
   plant                      FDA plant name and place, with FDA facility ID (FEI)
   latest_fda_classification  the plant's latest FDA inspection outcome on FDA's dashboard
                              (NAI = no action, VAI = voluntary action, OAI = official action)
   fda_fei                    FDA facility ID
   matched_by                 label = the drug's FDA label names the manufacturer and city;
                              city in 10-K = the annual report names the manufacturer and city
   source_section, source_text  where the link was read


HOW IT IS MADE
- Sources are public records only: SEC EDGAR filings, the FDA Data Dashboard, openFDA drug
  labels, FDA's FOIA reading room and warning-letter pages, EudraGMDP and MHRA. No personal data.
- Rules find candidate sentences in filings. A language model then labels each one (adverse or
  not, whose plant, what kind of site). In hand checks of kept sentences, 92-95 % were real
  inspection problems. This sample was also checked by hand, and 6 weak rows were removed.
- Every row keeps its public date, so the history can be replayed as it was known at the time.
  Plant IDs come from today's FDA plant list.

KNOWN LIMITS
- Many companies do not name their manufacturers publicly. Where a filing names a manufacturer
  but not the site, the full dataset lists every candidate site and marks it as such.
- FDA's dashboard lists final classifications only and leaves out pre-approval inspections.
  Company filings fill part of that gap, not all of it.

These files are records of what filings and regulators published, and when. They contain no
predictions, ratings or recommendations, and are not investment advice.
