A retrieval benchmark over the Epstein Files

Epstein Bench

Steve Bannon emailed Jeffrey Epstein eight words: “we r on the cusp of real power.”

He sent it at 10:22 on a Monday morning in February 2018. It sits somewhere in the roughly 3.5 million pages of records released by Congress and the Justice Department, between flight logs, scanned bank statements, and endless OCR sludge. We distilled the release into an 84,000-document haystack and asked the best AI systems in the world to find that email. The best one succeeds a little more than half the time.

This is the scoreboard.

54%

The current high score, set by the Claude Sonnet 5 agent at about $0.05 a question: 54% answered correctly with a valid citation. It gets the other 46% wrong.

Think yours does better? Submit a run.

Most retrieval benchmarks quiz AI on clean text. The Epstein Files are the opposite. This benchmark distills that release into an 84,000-document haystack of garbled scans, near-duplicate emails, and legalese, plus 1,038 questions a system can only answer by finding the right document and citing it.

From the files

Real questions from the benchmark. Each answer is a sentence hiding somewhere in the corpus, quoted exactly as the documents record it.

Hover or tap the black bars to unredact.

EFTA01820789
Who did Jeffrey Epstein ask to find him “the best codebreaker, NSA type”?
Buried in an email to a veteran TV journalist. The system has to surface the exact thread. “Can you find me the best codebreaker nsa type.”
EFTA00384983
What high-profile dinner was planned for September 20, 2013?
One line in a calendar: “TBD DINNER W/BILL GATES, TERJE, JAGBLAND, OTHERS (?)” The corpus holds thousands of entries like it. The trick is retrieving the right one.
EFTA01978481
What did Epstein say about Prince Andrew’s stay in New York?
A single sentence in a single email: “Prince Andrew will be staying with me for the week in New York.”
EFTA01522843
Which account was tied to Ghislaine Maxwell at J.P. Morgan?
A line item on a scanned bank statement that misspells her name: “CHISLAINE MAXWELL-HY ACCT” The answer key preserves the garble, because the garble is what is in the file.

The hardest questions ask a system to reconstruct the entire documented timeline of one person’s contacts with Epstein, scattered across dozens of files. Our best baseline scores near zero. the dossier tasks, 7 of 1,038

The scoreboard

full split, dataset v1.0, 1,038 tasks. Scored as cited answer correctness: the judged answer has to match the reference and cite a gold document. Higher is better.

#systemcorrect & cited95% CI$ / task
1Claude Sonnet 5 agentreference53.8%50.4 to 57.3$0.053
2Hybrid retrieval (BM25 + dense)reference49.0%45.4 to 53.3n/a
3Claude Opus 4.8 agentreference48.8%45.9 to 52.0$0.114
4BM25 keyword searchreference48.3%44.5 to 52.2n/a
5Dense embeddingsreference43.4%39.8 to 46.8n/a
Closed bookcontrol32.8%31.8 to 33.3n/a
Parametric probecontrol30.4%28.0 to 32.4n/a

The two controls answer from the question alone and never see the documents. Nearly all of their credit comes from correctly refusing the trap questions; on every question that requires the documents, they score zero. They sit on the board as a floor, unranked.

Full metrics: per-type scores, citation precision, retrieval recall, tokens
systemoverallmicrouncited single_hopaggregationtimelinedossier unanswerablefalse_premisecit. prec recall@5recall@20tok/task
agentic-sonnet-5 (reference)0.5380.5320.6070.5350.2860.4070.0001.0001.0000.4850.2500.32114,787
hybrid (reference)0.4900.4880.5410.4860.3130.2960.0320.8121.0000.4830.2720.594n/a
agentic-opus-4-8 (reference)0.4880.4600.5980.4630.1700.2960.0001.0001.0000.3680.2720.35719,586
bm25 (reference)0.4830.4750.5180.4760.2470.3330.0000.8441.0000.4620.2710.589n/a
dense (reference)0.4340.3930.4810.3830.2170.1850.0000.8440.9740.4060.2190.517n/a
closed_book (reference)0.3280.0660.3460.0000.0000.0000.0000.9691.0000.0000.0000.000n/a
parametric (reference)0.3040.0620.3460.0000.0000.0000.0000.8750.9470.0000.0000.000n/a

overall is the macro-average across task types; micro is the task-weighted average. The two diverge for the controls, which score only on the rejection types. uncited is correctness with the citation gate removed. false_premise rewards rejecting a fabricated meeting. Raw numbers in leaderboard.json.

How scoring works

An answer counts only if two things hold at once. A pinned LLM judge has to rule it equivalent to the reference, and at least one cited document has to belong to the pooled gold set. A right answer with no supporting citation scores zero. Confident prose gets no partial credit.

We also run the whole benchmark with the documents taken away. Given nothing but the question, a closed-book model answers under 2 percent of the single-hop questions correctly, and the model that wrote the questions manages about 5 percent from its training weights alone. That number is the contamination gauge. As future models train on this public release it will climb, and we will be tracking it.

The set also includes 38 questions built on a meeting that never happened. Every system we tested refuses them, so the trap itself is saturated. The difference shows up one level down: the agents name the specific fabricated claim, and nothing else can.

Every task survived a four-stage verification gauntlet before release. Full details in the methodology and dataset card.

Add your system

The rows above are reference baselines we built so a new system has something to measure against. Every score is recomputed from a submission’s raw predictions before it lands on the board.

Run your system on the full split, open a pull request with your predictions, and get on the board.

These are public records released by U.S. courts and Congress. Appearing in the files means appearing in someone's email, calendar, or financial records. It is not an accusation of wrongdoing. Epstein Bench measures whether AI can retrieve and cite what the documents say; it takes no position on any individual's conduct. For the avoidance of doubt, this neutrality does not extend to the crimes themselves.