Entity register · People · Redaction audit · entities.csv · people.csv · exclusions.csv

The DOJ Epstein production redacts in two directions at once

A reproducible redaction audit of the Department of Justice "Epstein Files" (EFTA) production. Public records only. Prepared by Publius. Every figure below is stated as-measured, with its denominator named; no victim or private-individual identity is recovered, reproduced, or asserted.

The finding, in one sentence

A single mechanical redaction process, applied by different hands to different material, fails in opposite directions: it over-redacts court-public substance and officials' names in government and court records, while it under-redacts private individuals' PII in the bulk email and financial material — and both failures show up in one reproducible random sample.

Method (reproducible)

• Frame. The full DOJ EFTA production, Bates EFTA00000001–EFTA02858497 (12 DataSets, 2,854,322 pages per the public index).
• Draw. A uniform random page-level sample over the Bates space, seed 44052026 (random.seed(44052026) then random.randint(1, 2858497), in order). Same seed reproduces the exact sample. This memo reports the first 120 draws.
• Coding. Each drawn page was retrieved from the public databases and every redaction coded to a fixed taxonomy: correct (a protected item properly removed), over (non-protected, court-public material withheld), under (protected PII left exposed). Two draws were not assessable; four did not resolve to any produced document (treated separately, below).

The numbers (n = 120, as-measured)

• 120 draws → 114 assessable pages (4 non-resolving, 2 not-assessable).
• Of the 114: 41 carried no redaction; 73 carried at least one.
• Of the 73 redaction-bearing pages: 60 correct · 5 under · 8 over.
• Over-redaction: 8 of 73 redaction-bearing pages (11%). Of those 8, 7 are substantive and every one is a court or DOJ record; the 8th is a trivial boilerplate vendor URL in an email.
• Under-redaction: 5 of 73 (7%) — every instance in the email/financial/image material.

The redaction, where assessable, is mostly correct (60 of 73) — which is what makes the failures on either end credible rather than cherry-picked.

Over-redaction: court-public substance and officials' names (Prong A)

All seven substantive over-redactions withhold material that is court-public, non-PII, or an official acting in an official capacity. Each is retrievable by Bates:

• The 2007 Non-Prosecution Agreement (EFTA00216124). A multi-line black box sits over the federal co-conspirator non-prosecution grant — the operative legal language of an agreement litigated to the Eleventh Circuit and public for years. The identical clause appears unredacted in the publicly released NPA, naming four potential co-conspirators granted non-prosecution — Sarah Kellen, Nadia Marcinkova, Adriana Ross, and Lesley Groff, none of whom was ever charged — so the box withholds court-public material, not a protected secret. The signing Assistant U.S. Attorney's name is boxed on all six signature pages of the instrument.
• The DOJ Office of Professional Responsibility report boxes the word "white" from the phrase "white collar" (EFTA02847341), and boxes a journalist's name in a footnote (EFTA02817216).
• Two civil filings box the identities of non-victim third parties — the intermediary who relayed messages between a bank CEO and Epstein (EFTA02817622), and a party in Giuffre v. Maxwell (EFTA02797781).
• A regulatory certificate boxes the Delaware Secretary of State's name (EFTA01694279); an FBI affidavit boxes a local detective's name (EFTA00227975).

Complementary pilot coding of the court-record-heavy DataSets found the same signature: a recurring "St. Thomas" venue/geography scrub, a redacted case caption, and the prosecuting AUSA's name — court- public material and official identities withheld.

Under-redaction: exposed private PII (Prong B — responsible disclosure)

Characterized systemically; no exposed identity is catalogued or reproduced here. In the same sample, the email/financial pass missed protected PII: a personal email address left clear where the adjacent phone number was boxed; a single traveler's name left clear in a booking list; an entire multi-thousand- line telephone toll record produced with the dialed numbers unredacted; and photographs "de-identified" by boxing the face only, which accepted practice treats as inadequate for an identifiable person. These are misses of protection, not over-withholding — the mirror image of the court-record failures. They are takedown candidates for the producing office, not material for publication.

A withholding signal (not yet a rate)

Four of the 120 draws (3.3%) resolve to no produced document in the public databases — candidate "assigned-but-withheld" gaps. This is the first such signal the sample has surfaced; smaller samples found only mid-document pages of long records. It is reported as a candidate figure only. Before any withholding rate is claimed it must be netted against benign causes — privilege holds, attachment/family boundaries, and re-production renumbering (dual EFTA_R1_ stamps confirm renumbering occurs).

Limits and discipline

• A random sample of 120; the point estimates carry a real confidence interval and are stated as in-sample rates, not exact production-wide percentages. Enlarging the seeded draw tightens them.
• Grand-jury and victim/witness material is handled at the systemic level only: redaction presence and correctness are recorded; no protected content is opened, transcribed, or reconstructed, and no boxed name is recovered from a redacted page — the NPA co-conspirator names above are cited from the separately and publicly released unredacted agreement, not extracted from this production's redaction.
• Every instance is a public document identified by Bates number and independently retrievable.

Why it matters

The government's own accountability record — the NPA it litigated, the OPR report on that NPA, and the filings about who enabled Epstein — is where this production most withholds court-public substance and the names of officials and non-victim actors. The same process, run over the bulk correspondence, instead lets private individuals' PII slip through. Both are failures of the same standard, and both are visible from a single reproducible draw of one-thousandth of one percent of the production.

Sample: seed 44052026, draws 1–120, frame EFTA00000001–EFTA02858497. Coded tables: coded-widened.csv (1–60), coded-widened-61-120.csv (61–120). Reproducible; corrections welcome against the Bates record.