About · Method
Methodology
How documents get into the archive, what the labels on a page mean, and how to tell what a page can and cannot support.
Source families and authority tiers
Every document in the archive belongs to a source family (court filing, agency report, inspector-general report, congressional material, company filing, press release, contract, dataset, arbitration record, news archive) and carries an authority tier. The tier governs what a document is allowed to support.
| Tier | What it is | Examples | What it can support |
|---|---|---|---|
| 1 | Primary official record | PACER and RECAP filings, SEC filings, the Federal Register, GAO, OIG, and PRAC reports, Treasury and SBA guidance, bills, committee reports, hearing records | Factual and legal-historical claims, when the metadata is clean |
| 2 | Official public communication | Agency and DOJ press releases, member releases, company press releases, testimony pages | What the issuer said or announced. Not treated as adjudicated truth |
| 3 | Structured public data | SBA loan-level files, USAspending, FOIA spreadsheets, EDGAR company facts | Quantitative claims, when the method is documented and the file version is named |
| 4 | Reputable secondary analysis | Academic papers, think-tank and watchdog reports, major investigations | Context. Never the sole support for a central factual claim when a primary record exists |
| 5 | Archive and helper material | Web captures, this site's own guides and derived tables | Navigation and analysis support only. Must point back to a higher tier |
Two consequences follow. A press release supports the statement that the agency announced something, not the truth of every sentence in it. An allegation is attributed to whoever made it unless a court or agency has adjudicated it, and the page says which.
Acquisition
Court records are acquired free-first. The order is: the archive's own holdings; then free sources, chiefly the RECAP archive at CourtListener, court and agency websites, and public document-sharing sites; then, last, a paid PACER purchase, made one document at a time only after both checks have failed. State-court records come from county clerk portals. Agency records come from the issuing agency's site, congressional records from Congress.gov, company filings from the filer's public disclosures. Every acquisition is logged with the source URL and the retrieval date.
Originals, checksums, and what they prove
Where a document page offers an original, the link reads "View original PDF" and states the file's size in bytes; the file is served as retrieved. Where the page says "Original unavailable", the document is published as extracted text only.
The data section goes one step further: every source file and every derived table in the release is listed with its SHA-256 checksum, byte size, and row count, and the release ships a SHA256SUMS file. To verify a download, compute its SHA-256 (on a Mac or Linux machine, shasum -a 256 filename; on Windows, certutil -hashfile filename SHA256) and compare it with the value on the page. A match means the copy has not changed in the archive's hands. It does not by itself prove the source served a faithful copy of what a court or agency holds; for that, the official copy linked from the page is the place to look.
Text, OCR, and originals
Text is extracted from the document's own text layer where one exists. Scanned documents go through Tesseract OCR. OCR text can contain recognition errors, dropped words, wrong line breaks, and flattened tables. Pages that fail quality checks are re-processed page by page. Where extraction has not produced usable text, the page says "Full text unavailable" rather than showing garbled output.
The original file controls for exact wording, pagination, signatures, exhibits, and legal effect. Everything else on the page, including the title and the record facts, is a navigation aid.
Dates
A court-filing page shows the filing date in its record facts, labeled "Filed". A source-document page shows the document's own date, labeled document_date. The date the archive retrieved or processed a file is tracked separately and is never presented as a filing date. Where a filing date cannot be established from the document or the docket, the field is left blank rather than guessed; those filings sit under "undated" addresses in their case.
How titles are derived, and how to tell
A page title comes from one of three places.
Verbatim. For court filings, the title is the docket text as the court entered it, followed by a dash, the case caption, and the docket number, for example "Minute Entry for proceedings held before Magistrate Judge Sharon A.… — USA v. RIVERA et al (Dkt. 12)". Long docket text is cut with an ellipsis rather than paraphrased. A verbatim title keeps the court's own wording, including its capitalization and abbreviations.
Composed. Where a document has no usable caption of its own, the title is assembled from the recorded fields in a fixed order: the document's type or subject, then the case caption or issuer, as in "Order on Motion to Expedite — 112 Genesee Street v. United States". A composed title never contains a phrase that does not appear in the document's recorded metadata.
Overridden by hand. Raw filing names are normalized by script with model assistance, and a human-maintained override table supplies the title where the script's result was wrong. Every override is checked against the document itself, its docket entry, or the heading on its first page, and the check is recorded with the override.
Some converted pages still carry a raw record identifier where a title should be. Those pages are meant to be held out of the index and the directories until a title is derived: as of September 12, 2026, 0 indexed pages still had a machine string for a title. If you see one, it is a bug: report it through the corrections page.
Descriptions and summaries
A source-document page has a one-line description built from its title and document type; a court-filing page that carries no summary has a one-line lede built from its record facts (the case, the filing date, the court, and how many filings from the case are in the archive). Neither is a reading of the document's contents. Where a document page carries a summary, a Claude model drafted it from the document's text on that page, a script checked every date, number, name and docket cite in the draft against that text, and a sample of every batch was read by a person before it appeared; the summary is labeled on the page. A summary is a navigation aid, never a source: cite the document. A summary says what the document is and what it contains; it does not adjudicate what the document proves. A page that carries one marks the summary block data-summary-state="model-drafted" and uses the summary for its description; a court-filing page with a summary carries no separate lede. A page whose text was too short or garbled to summarize carries none.
Datasets
Derived datasets ship as CSV with the official source files they were built from, the cleaning steps applied, and the known quirks of the underlying data. The headline program figures used across the site are the SBA's and USASpending's own: 11,468,210 PPP loans and $792.6 billion in current approved amounts; 3,680,124 COVID EIDL awards.
Known limitations
OCR errors survive in some extracted text. RECAP holds only the filings someone has bought and shared, so some dockets are incomplete here; each filing page says how many filings from its case the archive holds. Some documents have a local copy but no known original, and those pages say "Original unavailable". Some filings exist in the archive twice, from different sources; duplicates are kept and marked not-for-indexing rather than deleted. Title and date metadata carries a confidence level, and pages whose metadata is uncertain are held out of the site's index and sitemaps until reviewed: as of September 12, 2026, the court-filings section held 26,901 document pages, of which 12,204 are indexed and listed in the directory.
How to cite a page
Cite the document, then the archive: document title, issuer or court, date and date type, "archived at Pandemic Darlings," the page URL, and the date you retrieved it. Where a data page carries a SHA-256, include it. Each data page shows a suggested citation in this form; document pages do not yet.