How these pages are made

Each drug's FDA documents (the approval letters, the reviews and the label that FDA publishes on Drugs@FDA) are downloaded, fingerprinted with sha256 and read into sentences with their page numbers. Pages FDA stored as images, usually because they were redacted, are read by OCR and marked. The PDFs are not kept: every quote links to FDA's own copy.

The sentences are read into a knowledge graph by rules: trials, endpoints, doses, adverse events and the statistics attached to them. No language model is used. The trials, the drug's names and its approval status are drafted from the documents and the Drugs@FDA database, and NCT ids are matched on ClinicalTrials.gov only when the match is unambiguous.

A drug gets a page once automatic checks pass: FDA reviews are present, the pages have text, trials were found and the approval status was read. Drugs that fail a check wait for a person to look at them.

Every page is re-checked against FDA's copies weekly. When FDA changes or removes a document, the page says so.

What is not here: judgements. The pages quote FDA; they do not say whether a drug works or is safe for anyone.