A specific snapshot

The source CSV contains 205,958 change units across 27,535 documents and 82,856 document-page pairs. Those are the counts used here, rather than the report’s broader corpus totals.

2,190 AI visual inspections are provided upstream; 2,162 join to documents in this CSV. All 2,190 rows, including the 28 absent from this explorer, are retained in the Blackout index. Exact page matching finds 115 inspection records without a matching text-change page; these remain visible there too.

The report is dated March 5, 2026. The import date is October 21, 2026. Neither date establishes when a particular DOJ file changed.

Labels, with their limits

“AI visual inspection” means an upstream model inspected a rendered pair. It does not mean a person independently verified it. “Done” is a pipeline status, not a reliability score.

A source flag of visual_false_positive means the text difference was not treated as a substantive visible change by that pipeline. The report and CSV sometimes disagree; both should be checked against the actual PDFs.

Changes are different things

A file fingerprint detects changed bytes. Rendered-page comparisons detect visible differences. Extracted-text comparisons detect OCR and text-layer changes. These signals answer different questions.

A missing URL can mean a moved file, an access block, or a temporary failure. This snapshot makes no live deletion claims, and it does not establish why any change occurred.

Privacy before raw excerpts

Unreviewed free-text excerpts and AI narratives are withheld from the catalog because their upstream labels are not a reliable privacy review. Selected public pages include selectable source text, which can contain OCR errors.

Local PDF comparison runs entirely in the browser. It calculates SHA-256 fingerprints, renders selected pages, and highlights pixel differences. Misalignment, changed scans and page insertions can create differences; a colored region does not prove a redaction.