The DOJ's release under the Epstein Files Transparency Act runs to roughly 3.5 million pages. This oracle does not host the full corpus; it hosts a curated v0.1 subset — the named-case court records, the prior DOJ declassifications (Phase 1, Memos & Correspondence, Maxwell Proffer), and the DOJ Office of Inspector General report on Jeffrey Epstein's death in BOP custody — and pairs them with an encyclopedic editorial layer.
Ask the oracle a question. It cites primary documents by Bates number; the chips deep-link to the source PDFs at justice.gov.
The Epstein Oracle is a sister project to a family of editorial archives by Flex (Eurovision Oracle, World Cup Oracle, Annales Imperatorum, Conclave Oracle, Digital Art Oracle). Like them, it speaks from inside its subject: a research tool, not a neutral encyclopedia. Unlike them, the corpus is two orders of magnitude larger than any cache can hold, so v0.1 is a curated subset and future versions will add retrieval over the full ~3.5M-page release.
The oracle's voice is editorial: it cites primary documents for factual claims, distinguishes between direct quotation, multi-document inference, and outside reporting, and refuses to engage with claims that primary documents do not support. Where the DOJ has redacted a name, the oracle does not guess past the redaction. Where a name is attested in the corpus, the oracle will surface it with citation.
This is preservation work, not entertainment.