← The Election Integrity Files

Methodology & research integrity

How Atlas acquired, processed, and analyzed these documents — the sources, the tools, the models, the limits, and a live integrity scorecard. The objective is not to sound neutral; it is to make the reasoning auditable.

00

Research-integrity scorecard

Live figures from the database. Honest imperfect numbers, not a claim of perfection.

58/58
Files acquired
4/4
Source archives hashed
SHA-256 recorded below
9 / 49
Text-extracted / OCR
native vs. scanned-image documents
58/58
Evidence status assigned
58/58
Confidence rated
32 low · 23 moderate · 1 high · 2 not rated
4
Duplicate / variant versions
across 2 evidence families — counted once, not as independent corroboration
01

Acquisition

Retrieved
July 17, 2026 (source page dated July 17, 2026)
Archives downloaded
4 ZIP files → 58 PDF documents (269 pages)
Change detection
Source hashes recorded at retrieval (below); Atlas has not re-checked whether the White House has since altered the files.
Source archive SHA-256
Vulnerabilities-in-Electronic-Voting-and-Ballot-Counting-Systems.zip
165d85c466cb841643f7c5219340905889ae84cc5235264c78571cae22e7a607
Chinas-Acquisition-and-Exploitation-of-American-Voter-Data.zip
a823364d94ad5b9bca5dc42c10e8c4e0e782c3e6f0331e40dd600aaa27d041ce
Michigan-Voter-Registration-Investigation.zip
b73c4dee770bf6270174d3a55419498c7911ae133e0b01740bfac24ca69c40f6
Noncitizens-on-State-Voter-Rolls.zip
3291004b5f1631fba0f306a78027858b94d8ced1f740663c2fb6eccd7b34158e
02

Processing

Text extraction
Native text via pdftotext; scanned-image documents rasterized (200 dpi) and read with Tesseract OCR. OCR text is noisy and imperfect — the verbatim original is attached to every record.
Date normalization
Parsed from filename date stamps (e.g. 19AUG2020) and, where absent, inferred from the document’s own dated headers; precision recorded (day / month / year / range / unknown).
Clustering
The four investigative threads are the release’s own folders, not an Atlas judgment.
Duplicate detection
Documents with the same underlying filename appearing in more than one folder are linked as the same document and shown as one evidence family — five copies of one report count as one source, not five.
Relationships
“Branches” are computed deterministically: same-document duplicates, plus documents sharing a rare named entity. These are connections, not corroboration.
03

Analysis

Summaries & evidence cards
Generated by Claude (Anthropic) analyst passes, each reading the extracted text and instructed to attribute every claim, distinguish what a document says from what it establishes, and under-claim when uncertain. Generated from the documents — not from the source’s framing.
Ask the Files
Retrieves the most relevant documents (by entity/keyword), then answers using only those sources with inline citations, via the app’s language-model gateway (Kimi / Moonshot primary). It reports what the documents say; it does not adjudicate truth.
Confidence method
Two separate axes: (a) confidence the document is authentic, and (b) confidence its substantive claim is true. Conservative defaults — single-source, raw, or heavily redacted material is rated Low or Not rated.
Citation verification
Ask answers cite the specific documents retrieved; every record links to the original release archive. Page-level anchoring is not yet implemented (see limitations).
04

Limitations

  • Selection bias. The White House chose which documents to release, to support an announced position. This is not the complete government record and may omit contradictory material.
  • Redaction. Nearly every document is redacted; names, sources, agencies, and figures are withheld.
  • OCR quality. Scanned documents were machine-read; the extracted text contains errors, and some pages are barely legible.
  • Missing attachments. A few documents reference tables, spreadsheets, or exhibits that are not present in the released text (e.g. an FBI voter-fraud spreadsheet transmittal establishes only that a transmittal occurred).
  • Incomplete chains. Email threads and case files are partial; context before and after the released fragment is often absent.
  • No page-level citations yet. Citations resolve to the document, not the exact page or paragraph.
  • No independent-source layer yet. Atlas has not yet cross-checked these claims against court records, inspector-general findings, or independent reporting.
05

Who prepared this

Prepared by
Amplis Atlas — automated acquisition, extraction, and analysis pipeline.
Machine-extraction review
Passed (automated validation of file counts, hashes, and field completeness).
Editorial review
Automated, with operator spot-checks.
Legal review
Not conducted.
Subject-matter / external review
Not conducted. Outside methodology review (election-law, intelligence-analysis, and data-forensics) is invited.
06

Corrections & version history

Consequential changes are logged, not made silently.

Jul 17, 2026
Added per-document evidence cards: evidence-status label, two separate confidence axes (document authenticity vs. substantive claim), and a “what it says / establishes / does not establish” breakdown for all 58 records.
Jul 17, 2026
Added the corpus selection-bias disclosure and an “authenticated is not established” frame to the reading room.
Jul 17, 2026
Reworded the top summary from “traced to its source” to “linked to their original release copies” — Atlas links each file back to the White House archive, not to the underlying intelligence’s original evidentiary source.
Jul 17, 2026
Published this methodology page with source hashes and a live integrity scorecard.
Jul 17, 2026
Initial publication: timeline view and the Election Mail inbox view.