TLDR
Data Set 8 of the DOJ Epstein document release contains 419 MP4 surveillance videos totaling 412.5 hours of footage from the Metropolitan Correctional Center (U.S. Department of Justice [DOJ], 2025). VLM classification identified five camera types: corridor (159), cell tier (128), control room (106), common area (24), and interview room (2). All 419 videos have now been fully annotated, yielding 4,178 key frames across 368 videos with parseable wall-clock timestamps. Nearly all footage is 352x240 resolution at 4 frames per second, covering dates from March 1 through August 12, 2019 (PAPER TRAIL Project, 2026a).
What Data Set 8 Contains
The Epstein document release spans 12 data sets. Most contain PDFs, emails, and scanned documents. Data Set 8 is different. It contains 419 MP4 video files from the Metropolitan Correctional Center in Manhattan, the federal detention facility where Jeffrey Epstein was held from July 6, 2019 until his death on August 10, 2019 (DOJ, 2025).
The videos total 412.5 hours of footage. File sizes range from under a megabyte to over 23 megabytes per video. The metadata file creation dates cluster around September 2019 and late 2020, suggesting the videos were extracted from the facility's recording system in at least two batches (PAPER TRAIL Project, 2026a). The on-screen timestamps embedded in the footage itself tell a different story about when the cameras were actually recording.
The Camera Classes
A vision language model classified each video by examining a single representative frame and describing the camera's location and field of view (PAPER TRAIL Project, 2026a). All five camera classes have now been fully annotated with 4,178 key frames across 368 videos with parseable wall-clock timestamps — a 99.6% wall-clock parse rate. The 419 videos break into five distinct camera types:
| Camera Class | Videos | Frames | Guard Absent | Absent % |
|---|---|---|---|---|
| Corridor | 159 | 1,387 | 208 | 15.0% |
| Cell Tier | 128 | 1,583 | 92 | 5.8% |
| Control Room | 106 | 869 | 36 | 4.1% |
| Common Area | 24 | 243 | 35 | 14.4% |
| Interview Room | 2 | 96 | 7 | 7.3% |
| Total | 368 | 4,178 | 378 | 9.0% |
Corridor cameras capture hallways and movement between areas. Cell tier cameras show the rows of cells and adjacent common spaces. Control room cameras point at the monitoring stations themselves -- the desks where guards watch the other camera feeds. Common area cameras cover recreational and congregate spaces. Two videos capture an interview room, a small space with a desk, chairs, and a wall clock.
Of the 419 videos, 368 produced frames with parseable wall-clock timestamps, yielding 4,178 annotated frames total. The remaining 51 had no extractable frames or unparseable overlays. Corridor and cell tier cameras account for 68.5% of all videos, which is consistent with security priorities: movement and housing areas receive the most surveillance coverage (PAPER TRAIL Project, 2026a).
Resolution and Frame Rate
The technical specifications are remarkably consistent. Of the 419 videos, 413 share identical parameters: 352x240 resolution at 4 frames per second using the H.264 codec (PAPER TRAIL Project, 2026a). This is not high-definition surveillance. At 352x240, individual facial features are often indistinguishable. At 4 fps, motion appears choppy -- a person walking past takes perhaps 8-12 frames.
Six videos deviate from this standard. Four are 352x240 but recorded at 120 fps, suggesting a different recording mode. One is 854x480 at 29.97 fps, and one is 704x592 at 15 fps -- both significantly higher quality than the standard feed. The higher-resolution videos are among the most analytically valuable because they permit identification of individuals that the standard cameras cannot (PAPER TRAIL Project, 2026a).
What the resolution means in practice: you can distinguish a guard in uniform from an inmate in a jumpsuit from a person in civilian clothing. You can see whether doors are open or closed. You can count the number of people in a frame. You generally cannot read name tags, identify specific individuals by face, or determine what someone is carrying unless it is large enough to occupy multiple pixels.
The Temporal Window
On-screen timestamps embedded in the video footage span two distinct periods. The first window covers March 1 through March 12, 2019, with 54 date references. The second and larger window runs from July 5 through August 12, 2019, with 364 date references (PAPER TRAIL Project, 2026a).
The July-August window is the critical period. Jeffrey Epstein was arrested on July 6, 2019 and found dead on August 10, 2019 (DOJ, 2025). The surveillance footage covers nearly the entire duration of his final detention, from the day after his arrest through two days after his death. The March footage predates his arrest by four months, capturing what the facility looked like during routine operations.
This temporal distribution means the corpus includes both a baseline (March) and the period of interest (July-August), enabling comparison of staffing patterns, movement frequency, and surveillance coverage between normal operations and the period surrounding Epstein's detention and death.
What Was Not Released
The 419 videos capture only a fraction of MCC's camera infrastructure. A federal detention facility of this size operates dozens of cameras simultaneously, 24 hours a day. The released footage represents specific cameras on specific dates -- not continuous coverage of the entire facility (DOJ, 2025).
Several categories of footage are absent. There is no audio in any video. There are no exterior cameras showing entrances, exits, or the perimeter. There are no cameras showing booking areas, medical facilities, or administrative offices. The coverage concentrates on housing units, corridors, control rooms, and a handful of common spaces (PAPER TRAIL Project, 2026a).
The control room footage is particularly noteworthy for what it implies. These 106 videos show the monitoring stations where guards are supposed to be watching the other cameras. All 106 have now been classified and 55 produced frames with parseable wall-clock timestamps, yielding 869 annotated frames — the full results are detailed in What the Control Room Revealed. If a control room camera shows an empty desk, it means nobody was watching the feeds from the corridor and cell tier cameras during that period. The monitors watching the monitors are themselves being monitored. Cross-camera correlation shows that 98% of floor-level guard gaps have no overlapping control room footage, making it impossible to determine whether electronic monitoring was maintained during physical guard absences.
The 278 guard gaps identified across the full corpus tell a story of escalation: absence rates climbed from a July baseline of 4% to a peak of 34.2% on August 2, then snapped back to baseline within 48 hours of Epstein's death. The pattern is phase-driven, not event-driven — the failure began after suicide watch was removed. The arrest-night concurrent gap on July 6, 2019 is the only moment in the corpus where both the floor and the control room simultaneously showed zero guards present.
These surveillance findings form the evidentiary foundation for assessing congressional oversight of the DOJ's compliance with the Epstein Files Transparency Act.
References
PAPER TRAIL Project. (2026a). Jail surveillance video analysis results [Data set]. Script 34 output, _exports/video_analysis/.
U.S. Department of Justice. (2025). Epstein document release, Data Set 8: MCC surveillance footage [Government records].
This investigation is part of the SubThesis accountability journalism network.