FILE ANALYSIS & FINDINGS

1. The Scale of the Release

The DOJ released Jeffrey Epstein’s files across 12 separate ZIP/TAR archives totaling approximately 461 GB of compressed data, expanding to over 1.86 TB on disk. The release contains 904,395 individual files, of which approximately 900,956 are PDFs.

To put this in perspective: if you spent 30 seconds reviewing each document, working 8 hours a day with no breaks, it would take 2.6 years to get through all of them. The format choices made by the releasing agency ensure this is exactly what would be required for any meaningful review.

2. Format & Delivery Choices

Everything Is a PDF

Nearly every file in the release — 99.6% — is a PDF. This includes emails, text messages, screenshots, photographs, handwritten notes, spreadsheets, legal filings, and blank pages. Each individual email is its own separate PDF file. A conversation thread that might be a single scrollable email in your inbox becomes 5–15 individual PDF files that must be opened, read, and cross-referenced manually.

Why Not Just Release the Native Formats?

The emails in these files originally existed as .msg, .eml, or in PST archive format. The images existed as .jpg and .png files. The spreadsheets were .xls or .csv. Converting everything to individual PDF files:

• Strips all metadata (timestamps, sender info, thread context)
• Makes full-text search impossible without OCR processing first
• Inflates the total file count from what would be ~50,000 logical documents to 900,000+ individual PDFs (containing over 3 million actual pages)
• Forces every researcher to independently solve the same extraction problems
• Fragments email conversations across thousands of unlinked files

Images Stored as PDFs

Photographs, screenshots, scanned documents, and even business cards are all wrapped in PDF containers. A simple JPEG photograph that would be 200KB becomes a 1.5MB PDF. Our extraction pipeline pulled 835,718 images out of these PDF wrappers. Of those, over 564,000 appeared blank at the pixel level. However, our Deep Sweep later recovered hidden text from 400,913 of those pages — they were blank as images but contained embedded text layers that yielded 40.3 million additional words.

Concordance / IPRO Load File Format

Dataset 10 (302,600 files) uses Concordance/IPRO litigation support format. This is an industry tool used by law firms for document review — it assigns each page a Bates number and stores it as a single-page TIFF or PDF. A 47-page deposition becomes 47 individual files named EPSTEIN-00384721.pdf through EPSTEIN-00384767.pdf. Reassembling documents from this format requires the accompanying load file (DAT/OPT), which is not always included or is itself in a proprietary format.

3. Content Quality Breakdown

We performed a systematic analysis of the file contents by manually categorizing a 500-document sample from Dataset 11 (the largest dataset at 331,997 files). Because these files are released in numbered batches, we reviewed contiguous ID windows (plus spot checks outside those windows) so adjacent files could preserve missing thread context. Results:

What “Admin / Scheduling” Means (43%)

Nearly half the files are scheduling emails, appointment confirmations, flight arrangements, and logistical coordination. These are emails like: “Haircut at 2pm. Vegan lunch with Mort Zuckerman at 12:30. Car to Teterboro at 4.” While even scheduling details can be evidentiary (they prove who was where, when), the volume of mundane logistics dramatically dilutes the substantive material.

What “Substantive” Means (38%)

The genuinely significant material includes financial records, legal correspondence, witness depositions, wire transfer records, property documents, communications between Epstein and high-profile associates, and documents from criminal proceedings. This includes:

• Wire transfer records and bank statements from multiple financial institutions
• Correspondence with attorneys (Dershowitz, others) regarding legal strategy
• Property records for homes in New York, Palm Beach, New Mexico, USVI
• Flight manifests and travel arrangements referencing specific individuals
• Communications between Ghislaine Maxwell and Epstein
• Documents from the 2006–2008 Florida investigation and plea deal
• Meeting schedules that place specific associates at specific locations
• Leon Black / Apollo Global financial relationship documentation
• MCC Metropolitan Correctional Center records (Dataset 8)

What “Noise” Means (18%)

Nearly one-fifth of the files are genuinely useless for investigative purposes: blank pages, cover sheets, fax headers, duplicate transmissions, auto-forwarded news articles, and personal messages with no evidentiary value (e.g., “Miss you! Good night xxxxoo” texts between Epstein and unnamed contacts).

Document Type Distribution

Across all datasets, approximately 96% of files are emails or email-derived content (email bodies, attachments, forwarded messages). The average document contains approximately 1,492 characters of extractable text — about half a printed page.

4. Image Extraction & Classification

Because every document is a PDF — including photographs and screenshots — we built an image extraction pipeline that pulls embedded images out of every PDF and classifies them using CLIP (Contrastive Language–Image Pre-training), an AI vision model. This allows us to categorize the visual contents of the release without opening each PDF manually.

Top Categories by Image Count

Images classified by CLIP into descriptive categories. The top two categories alone account for 68.5% of all images.

Full Category Inventory

5. Blank Pages & Waste

A significant portion of the release consists of completely blank pages — white rectangles stored as full PDF documents with filenames, file system entries, and archive space. These are not redacted pages (which would typically show redaction bars or stamps) — they are genuinely empty pages.

Blank Page Breakdown by Source

6. Dataset-by-Dataset Breakdown

The 12 datasets vary enormously in size and content. Three datasets (DS9, DS10, DS11) contain 98% of all files. The remaining nine datasets combined hold fewer than 16,000 files.

7. What It Takes to Process These Files

There is no “download and read” path for these files. Any meaningful analysis requires building custom software infrastructure. Here is what our pipeline involves:

Technical Requirements

Software Stack

• PyMuPDF (fitz) — PDF text and image extraction
• OpenAI CLIP — Zero-shot image classification (ViT-B/32)
• Pillow + NumPy — Image analysis, blank detection (mean pixel > 252, std < 3)
• Tesseract OCR — Optical character recognition for scanned pages
• Python 3.11 — All scripts, batch processing, report generation
• Static HTML/JS — Website with search, maps, timeline, cast tracking

8. Editorial: The Document Dump Strategy

The Pattern

Document dumps share common characteristics, and this release exhibits all of them:

• Volume as obstruction: 900,000 documents (3 million+ pages) sounds like full transparency. In practice, it means no single human can review the material. The sheer volume is itself a form of concealment.
• Hostile formatting: Converting native email archives to individual PDFs destroys searchability, threading, metadata, and context. This is a choice — the original files could have been released in their native format.
• Noise inflation: Including 564,000+ blank pages, duplicate transmissions, and junk inflates the file count and dilutes substantive content. Every blank page a researcher opens is time not spent on real evidence.
• Fragmentation: Splitting material across 12 separate archives with inconsistent naming schemes, mixed formats, and no cross-referencing index means researchers must independently solve the same organizational problems before any analysis can begin.
• No index or manifest: A 900,000-document, 3-million-page release with no master index, no date range guide, no subject categorization, and no document-type breakdown. The releasing agency knows what these files contain — they reviewed them before release. The choice not to include an index is deliberate.

The Effect

The practical result is that meaningful analysis requires:

• 2+ TB of storage space
• Custom software engineering capability
• Machine learning infrastructure for classification
• Weeks to months of processing time
• Domain expertise in legal, financial, and criminal investigation documents

This effectively limits genuine analysis to well-funded organizations, while allowing officials to point to the release as evidence of transparency. The files are “public” in the same way that a needle in 3 million pages is “findable.”

What Genuine Transparency Would Look Like

• Release emails in their native format (PST/EML) with intact threading and metadata
• Provide a searchable index with document types, date ranges, and subject categories
• Remove blank pages and duplicates before release
• Merge single-page Concordance/IPRO files back into complete documents
• Release images as image files, not PDF-wrapped containers
• Provide a single consolidated archive rather than 12 fragmented datasets
• Include a document manifest mapping Bates numbers to logical document boundaries

9. Content Samples

To illustrate the range of material, here are representative samples from different content categories, drawn from actual files in the release.

Scheduling / Admin (43% of content)

Substantive Financial (38% of content)

Personal / Noise (18% of content)

MCC Prison Records (Dataset 8)

Legal Correspondence

THE PDF FILES

Analysis based on processing 904,395 files across 12 DOJ datasets.
Image classification via CLIP (ViT-B/32). Text extraction via PyMuPDF.
All scripts and methodology are open source and reproducible.

← Back to Main Site  ·  🕸️ The Inner Circle  ·  ⚖️ Complicity Gradient  ·  💣 Explosive Documents  ·  🔍 Deep Sweep  ·  Cast Hub  ·  Search Files