Every query is sent to public archive APIs. Titles, authors, dates, repositories and URLs are copied from the archive's response. A model plans the searches and ranks the results; it does not supply records.
Quotes are shown only where the text has been matched against the archive's own copy. Where that check cannot be made, no quote is shown.
Every exported bibliography ends with a receipt: a way to re-check each record against the live archive, so the claim can be tested rather than trusted.