TLDR
VLM frame-by-frame annotation of 4,178 key frames across all 368 surveillance videos with parseable timestamps reveals guards present in 91.0% of frames and absent in 9.0%. The 378 guard-absent frames form 278 discrete gaps totaling 202.3 minutes of unmonitored time. The longest single gap lasted 45 minutes 42 seconds on a common area camera dated March 12, 2019 (PAPER TRAIL Project, 2026a).
How Key Frames Were Selected
Not every frame in 412.5 hours of footage can be individually annotated. The extraction pipeline uses two complementary methods to identify the frames that matter most. First, scene-change detection (threshold 0.03) identifies moments where the visual content shifts meaningfully -- a door opening, a person entering, lighting changing. Second, fixed 60-second interval sampling ensures that even static footage is periodically checked. After deduplication, this process yielded 4,178 key frames from 368 videos, capped at 120 frames per video (PAPER TRAIL Project, 2026a).
Unlike the initial analysis that covered only corridor (159) and cell tier (128) cameras, all five camera classes have now been fully annotated: corridor, cell tier, control room (55 with timestamps of 106 classified), common area (24), and interview room (2). This provides a complete picture of guard presence across every camera type in the corpus.
The Guard Presence Baseline
Each of the 4,178 key frames was processed by a vision language model that answered a structured set of questions: Is a guard present? How many guards? What is the door status? How many people are visible? What clothing are they wearing? (PAPER TRAIL Project, 2026a).
The results:
| Metric | Value |
|---|---|
| Frames with guard present | 3,800 (91.0%) |
| Frames with guard absent | 378 (9.0%) |
| Total guard absence gaps | 278 |
| Total gap duration | 12,139 seconds (202.3 minutes) |
| Average gap duration | 44 seconds |
| Timeline segments | 686 (408 present, 278 absent) |
A 91.0% guard presence rate sounds high. But the 9.0% absence rate, when mapped to continuous time, produces 202.3 minutes of cumulative unmonitored coverage across 368 cameras -- more than three hours of total unmonitored time. The gaps are not distributed evenly -- they concentrate in a handful of videos with extended absences (PAPER TRAIL Project, 2026a).
Guard Absence by Camera Class
The expanded annotation reveals significant variation in guard absence rates across camera types:
| Camera Class | Videos | Frames | Guard Absent | Absent % |
|---|---|---|---|---|
| Corridor | 159 | 1,387 | 208 | 15.0% |
| Common Area | 24 | 243 | 35 | 14.4% |
| Interview Room | 2 | 96 | 7 | 7.3% |
| Cell Tier | 128 | 1,583 | 92 | 5.8% |
| Control Room | 55 | 869 | 36 | 4.1% |
Corridors and common areas show the highest absence rates (15.0% and 14.4%), while the control room shows the lowest (4.1%). The control room's relatively low absence rate is expected -- monitoring stations should be staffed. But even there, 36 frames across 55 videos show no guard present (PAPER TRAIL Project, 2026a).
The Ten Longest Gaps
| Rank | Video | Duration | Camera Class | On-Screen Date |
|---|---|---|---|---|
| 1 | EFTA00033029 | 45 min 42 sec | Common Area | 03/12/2019 |
| 2 | EFTA00033122 | 43 min 08 sec | Common Area | 07/23/2019 |
| 3 | EFTA00033037 | 41 min 00 sec | Cell Tier | 07/05/2019 |
| 4 | EFTA00033213 | 20 min 57 sec | Corridor | 08/08/2019 |
| 5 | EFTA00033236 | 15 min 19 sec | Control Room | 07/06/2019 |
| 6 | EFTA00033211 | 15 min 00 sec | Cell Tier | 08/08/2019 |
| 7 | EFTA00033123 | 7 min 52 sec | Corridor | 07/23/2019 |
| 8 | EFTA00033058 | 3 min 00 sec | Corridor | 07/09/2019 |
| 9 | EFTA00033060 | 2 min 00 sec | Corridor | 07/10/2019 |
| 10 | EFTA00033126 | 1 min 41 sec | Common Area | 07/24/2019 |
The longest gap -- 45 minutes 42 seconds -- occurs on EFTA00033029, a common area camera with an on-screen timestamp of March 12, 2019 (PAPER TRAIL Project, 2026a). This predates Epstein's arrest by nearly four months, suggesting extended guard absence was not unique to the period of Epstein's detention.
The second longest gap (43 minutes 8 seconds) falls on July 23, 2019 -- the same day as the reported suicide attempt. Gaps #4 and #6 both fall on August 8, 2019 -- the day Epstein signed his last will. Gap #5 is the longest control room gap (15 minutes 19 seconds) and falls on July 6, 2019 -- the day of Epstein's arrest (PAPER TRAIL Project, 2026a).
What Clothing Tells Us
The VLM annotates each frame with descriptions that include clothing type. This enables a rough classification of who appears in each frame:
- Guard uniforms: Dark clothing consistent with Bureau of Prisons uniform standards. The primary indicator of staff presence.
- Inmate jumpsuits: Orange or tan one-piece garments. The standard inmate attire at MCC.
- Civilian clothing: Street clothes -- shirts, pants, jackets -- that match neither guard uniforms nor inmate jumpsuits.
Of the 42 unguarded door status changes, frames describe inmates, individuals in civilian clothing, or empty corridors. The presence of civilian-clothed individuals during guard absence periods raises questions about who had access to secure areas without escort. Civilian clothing in a federal jail is typically associated with attorneys, investigators, medical professionals, or facility administrators -- all of whom ordinarily require escort in housing areas (PAPER TRAIL Project, 2026a).
What 9.0% Means
The 9.0% guard absence rate is a floor estimate, not a ceiling. The frame sampling method captures snapshots at scene changes and 60-second intervals. A guard who leaves for 30 seconds between two sampled frames would not register as absent. The true absence rate could be higher than what the sampled frames reveal (PAPER TRAIL Project, 2026a).
The 202.3 cumulative minutes also understates the operational impact because the gaps are not randomly distributed. They cluster in specific videos, meaning specific locations experienced extended periods without oversight. A single 45-minute gap on one camera is operationally different from 45 one-minute gaps across 45 cameras. The former means one location was completely unmonitored for nearly three-quarters of an hour.
These findings describe what the cameras recorded. They do not establish why guards were absent, whether absences violated specific BOP policies, or whether the gaps contributed to any particular outcome. What they establish is a measurable baseline: across all 368 analyzed surveillance videos from a federal detention facility, guards were not visible in 9.0% of sampled frames, forming 278 gaps totaling over three hours of unmonitored time.
The per-date breakdown reveals a pattern of escalation from 4% to 34% concentrated in the 12 days between suicide watch removal and death. The Lag Effect analysis shows the failure was phase-driven, not event-driven. During 42 of these gaps, doors changed state without guard presence. And for 98% of floor gaps, no overlapping control room footage exists to determine whether electronic monitoring was maintained.
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.