SEC Account Briefs

A decision-ready account brief for any US filer — every sentence cited to an SEC filing.

Give it a company. Get back a structured brief: what changed recently, who came and went at the top, what the financial signals are, what to watch — each point tied to a specific 8-K item with a link to the filing on sec.gov. Run it with no AI at all (every line is a cited fact), with your own Anthropic key, or in managed mode where a bounded model writes two to four sentences per section and a validator throws out any sentence it can't cite.

Who it's for

• Sales teams doing pre-call research on enterprise accounts: the brief is the first page of your account plan, and every claim in it survives a "where did you read that?"
• Account-research and investment-research teams who need a repeatable, citable summary per company rather than a fresh reading of every filing.
• Anyone who has been burned by an AI summary that made something up. This one structurally cannot — see How the citation guarantee works.

What you get

One dataset item per company:

Part	What it holds
company	Name, ticker, CIK — as SEC names the registrant
window	The evidence window (from, to, lookbackDays)
coverage	How many filings the brief draws on, by form, with newest and oldest dates and cache freshness
sections[]	Four synthesized sections — Recent material events, Management and governance, Operating and financial signals, Risks and watch items — each a list of sentences with citations (E1, E2, …)
citations[]	Every cited evidence item: filing date, form, headline, filingUrl, primaryDocumentUrl
limitations[]	The window, caps hit, sections with no evidence, any section that fell back to extractive text and why, and the model cost if a cap was exceeded
evidenceSet[]	The exact events the brief was built from, with retrievedAt and classifierVersion — regenerate the same brief from the same evidence
model	null in extractive mode; otherwise the mode, model id, and effort
costs	Mode, model, tokens, estimated USD, cap, and an exceeded flag — BYOK and managed costs are never mixed

Each section says where its text came from: synthesized (model sentences that passed validation), extractive (deterministic cited lines), or empty (no evidence in the window — stated, never filled).

Sample section (extractive mode, real run)

{
  "id": "financial",
  "title": "Operating and financial signals",
  "source": "extractive",
  "sentences": [
    {"text": "2026-08-26 · 8-K Item 2.02 · Financial update — On August 26, 2026, NVIDIA Corporation, or the Company, issued a press release announcing its results for the quarter ended July 26, 2026.", "citations": ["E16"]},
    {"text": "2026-05-20 · 8-K Item 2.02 · Financial update — On May 20, 2026, NVIDIA Corporation, or the Company, issued a press release announcing its results for the quarter ended April 26, 2026.", "citations": ["E18"]}
  ]
}

How the citation guarantee works

1. Evidence first. The brief's skeleton is built from the company's classified SEC events (from the companion SEC Event Intelligence cache). Each event becomes a numbered evidence item — E1, E2, … — with its quoted filing text and links.
2. The extractive brief is complete on its own. Every line is a dated, cited fact. In extractive mode this is the whole brief, and no model runs.
3. Synthesis is bounded. In byok or managed mode, one structured model call writes two to four sentences per section, using only the evidence it was given, and must tag every sentence with the evidence ids it rests on.
4. The validator decides. A sentence with no citation, a citation to an id that doesn't exist, or a citation to evidence from a different section rejects the whole section. A rejected section falls back to its extractive text, and limitations says why.

The model cannot add a fact, because a fact without a citation never reaches the output. This is enforced in code, not in a prompt.

Input

{
  "companies": [{ "cik": "0000320193", "ticker": "AAPL" }],
  "asOfDate": "2026-09-12",
  "lookbackDays": 365,
  "sections": ["material-events", "governance", "financial", "risks"],
  "detail": "standard",
  "aiMode": "extractive"
}

Field	Notes
companies	1–10 companies by SEC CIK. Ticker and legal name are display hints.
asOfDate, lookbackDays	The evidence window: lookbackDays (1–1095) ending at asOfDate (default today).
sections	Which synthesized sections to include. Coverage, citations, and limitations are always included.
detail	brief (up to 5 events per section) or standard (up to 12).
aiMode	extractive — no model. byok — your Anthropic API key, passed as a secret input. managed — the operator's model.
anthropicApiKey	BYOK only. Stored as a secret; never written to the dataset or logs.
model	BYOK only: claude-opus-5 (default), claude-sonnet-5, or claude-haiku-4-5.
maxModelCostCents	Managed only: 1–20 cents per brief. A brief that exceeds it is still delivered, and flagged.

Pricing

Pay per brief, charged only for a completed brief:

Mode	Event	Price	Model cost
Extractive	byokBrief	US$0.35	none
BYOK	byokBrief	US$0.35	on your key; the brief reports estimated tokens and USD at list price
Managed	managedBrief	US$0.50	included, capped at US$0.20 per brief

Prepaid volume blocks

For teams buying ahead, prepaid blocks are consumed in 500-brief increments. These are commercial bundle prices; the Actor's Apify pay-per-event event prices remain US$0.35 for extractive/BYOK and US$0.50 for managed.

Briefs	BYOK	Managed
500	US$175 (US$0.35/brief)	US$250 (US$0.50/brief)
1,000	US$325 (US$0.325/brief)	US$475 (US$0.475/brief)
2,500	US$750 (US$0.30/brief)	US$1,100 (US$0.44/brief)
5,000	US$1,400 (US$0.28/brief)	US$2,000 (US$0.40/brief)

Bundle pricing is separate from the Actor's fixed event prices and does not change Apify pay-per-event billing.

A company with no events in the window still yields a (mostly empty) brief; it is charged as a completed brief because the coverage, citations, and limitations sections are real answers.

If the run reaches the spend limit you set when starting it, the dataset holds only the briefs you were charged for, and the run summary reports charging.limitReached: true with the counts.

Limits — please read

• Coverage follows the SEC Event Intelligence cache. A brief reports freshness.lastCheckedAt and freshness.stale; when it is stale or the company has never been checked, run SEC Event Intelligence for that company first — the brief itself never fetches from SEC.
• Evidence is the 8-K item section, not the exhibits. Earnings numbers, for example, live in the press-release exhibit; the brief will tell you results were announced and link the filing, not restate the figures.
• Sections with no evidence stay empty. The brief never fills a gap with outside knowledge.
• In managed mode, sentences are the model's words — but every one is cited. Read the citations; the evidenceSet lets you regenerate the brief from exactly the same inputs.
• Model cost is estimated from list prices, reported separately for BYOK and managed, and never blended with the brief price.
• No advice, no predictions. Sentences describe what filings say.

Compared with alternatives

	Reading the filings	A generic AI summary	SEC Account Briefs
Time per account	30–60 minutes	seconds	seconds
Can you verify a claim?	yes, by hand	usually not	every sentence carries citations → primaryDocumentUrl
Can it invent a fact?	no	yes	no — uncited sentences are rejected in code
Repeatable	no	no	evidenceSet + model metadata on every brief

FAQ

Which mode should I start with? extractive. It is free of model cost, every line is a fact, and it shows you exactly what the brief would be built from. Switch to managed when you want prose.

What's the difference between BYOK and managed? Same brief structure and the same validator. BYOK uses your Anthropic key and you pay Anthropic directly; managed uses ours with a hard cost cap per brief.

What happens if the model refuses or errors? The brief is still delivered, fully extractive, with the reason in limitations. The run does not fail.

Can it brief any company? Yes, once SEC Event Intelligence has fetched it; the brief tells you when it has not.

Where do the events come from? The companion Actor, SEC Event Intelligence — same cache, same classifier, same evidence.

Support

Use the Actor's Issues tab. For a wrong or unsupported sentence in managed mode, include the eventIds from evidenceSet — the brief is reproducible from them.

Changelog

• 0.1 — Private build. Extractive, BYOK, and managed modes; citation validator with extractive fallback; per-brief model-cost cap.