edgarwiki

Methodology

What the numbers on this site mean, how they are produced, and what they cannot support. Read this before quoting anything here.

Sources

Everything comes from the SEC's EDGAR system. Two form types matter: UPLOAD, the staff's comment letter to a company, and CORRESP, the company's reply. This build is based on CORRESP, because a response letter conventionally reproduces each staff comment verbatim before answering it — putting the comment, the response, and the outcome in one document.

Extraction

Comments and responses are separated by deterministic parsing — regular expressions anchored on response headings and numbered comment openers. No language model produces any text, number, name or date on this site. Each stored comment keeps its character offsets into the cleaned source document, and a verifier re-reads every source and asserts that the text at those offsets still matches what is stored. That check currently passes on every quotation in the corpus.

Counting

Filing counts come from EDGAR's quarterly form.idx index files. They are not taken from EDGAR full-text search, which caps result totals at 10,000 and reports the cap as though it were a count — a trap that silently corrupts any long-run series built on it.

What is in the corpus right now

Measured on the corpus as it stands: 4,297 staff comments drawn from 552 substantive comment-response letters (1,348 CORRESP filings examined, 2023-01-03 to 2025-12-17).

Letter kindFilingsShare
substantive55240.9%
acceleration_request49436.6%
short_cover_letter14710.9%
unclassified14210.5%
supplemental_materials60.4%
transmittal_letter50.4%
unlabelled10.1%
withdrawal10.1%

Most CORRESP filings are not comment responses at all — a large share are Rule 461 acceleration requests and similar procedural cover letters containing no staff comments. They are identified and excluded rather than counted, which is why the substantive figure is far below the raw form count.

Topic matching

Issue pages are assembled by matching regular expressions against the verbatim comment text. This is transparent and reproducible, and it is coarse: it will group a passing mention with a substantive objection, and a comment raising several issues appears under each. Span-justified classification is the next piece of work. Until it lands, treat topic counts as an index, not a statistic.

What these numbers cannot support

Corpus built 30 August 2026.