How we read filings, so you don’t have to read them all

• 10-K and 10-Q — annual and quarterly filings, the heaviest reading by far. We focus on MD&A, Risk Factors, and the notes to the financial statements.
• 8-K — material event filings. Earnings, executive departures, restatements, agreements, capital raises.
• Live EDGAR feed — the moment a filing appears on SEC.gov, our pipeline picks it up. The “LIVE” badge on a finding means it landed via live ingest rather than the slower bulk archive.

For each filing, we extract the XBRL TextBlock for every disclosure section the company chose to publish. The current quarter’s text is diffed against the prior quarter’s — both word-by-word and section-by-section. A finding is created when the change crosses an editorial threshold: a new sentence in Risk Factors, a material rewrite of MD&A, a removed disclosure that previously appeared. Mechanical filters are the first pass; a Claude call does the second — producing the plain-English “why it matters” line and a severity label.

• Critical — material disclosure significance: a risk added that the company didn’t previously flag, a going-concern doubt, a restatement, a counterparty failure.
• Warn — a significant rewrite without a clear directional valence; worth a look but not necessarily acute.
• Info — a noticeable language change that isn’t materially severe; included for completeness.
• Positive — a previously-disclosed risk removed, or new optimism in MD&A that wasn’t there before.

Severity describes disclosure significance, not investment quality. A “critical” finding is not a sell signal; an “emerald” finding is not a buy signal. Drift is a publication, not an adviser.

Every claim on a finding card links to the source paragraph in the original filing. Click any citation pill [1] in a chat answer, or click a finding’s “Source” tab — both deep-link to SEC EDGAR. If a fact isn’t in the filing, Drift doesn’t claim it.

• Tell you whether to buy, sell, or hold any security.
• Predict future stock prices, returns, or performance.
• Tailor content to your personal portfolio. The brief is the same for every Pro subscriber.
• Aggregate or display other users’ predictions, journal entries, or watchlists. Your data is yours.
• Make accuracy claims about user predictions or about Drift’s severity calls.

These commitments are how Drift operates within the publisher exemption of the Investment Advisers Act of 1940 (Section 202(a)(11)(D), Lowe v. SEC, 1985). Same legal foundation as Substack newsletters and the financial media.

The Friday digest is the heart of the product. Once a week, we surface the most material drift across your companies in one short email. The web app is for the deeper dive — the paragraph-level diff, the contemporaneous market context, your private journal.

We deliberately don’t do a real-time firehose. Real-time isn’t a wedge worth fighting for; weekly editorial discipline is.

Most market datasets quietly drop companies that die — so every statistic computed on them is biased toward survivors. Ours is built the other way around. Our outcome universe tracks 10,144 companies that hit a terminal outcome — 1,476 delisted or bankrupt, 8,668 that simply stopped filing (“went dark”) — and they stay in every base rate we publish. When the Precedent panel says “of 660 companies that disclosed a going concern, 2% were eliminated within 24 months,” the dead companies are in both the numerator and the denominator.

Three guardrails apply to every outcome statistic we show:

• Thin cohorts are suppressed. Fewer than 5 comparable companies → we say “not enough history” instead of showing a percentage.
• Industry fallbacks are labeled. If your company’s industry cohort is too thin and we fall back to the all-industry base rate, the panel says so.
• History, not a prediction — printed in the statistic itself, because base rates describe the past; they do not forecast any specific company.

Our findings come from automated detection plus editorial review, and automated detection has a false-positive rate — some flagged changes turn out to be routine boilerplate. We’d rather you know that than discover it. Three things keep it honest:

• Every finding carries its verbatim source quote and a deep-link to the filing paragraph — checking us takes one click, not a subscription to our judgment.
• The Friday digest is editorially reviewed— what reaches your inbox passed a human filter, not just a model.
• We publish corrections (see below) and continuously measure detector precision against adjudicated samples; suppression rules for known false-positive classes (e.g. routine annual-meeting votes) ship as we find them.

We will. The corpus is huge, language is ambiguous, and automated mining is imperfect. If you find a finding that misreads a filing, email corrections@driftbrief.com with the finding URL — we read every report and correct or retract within 24 hours when warranted.