How it works, and what it does not do.

In one paragraph: we collect public filings and official public data. We store the exact source response, extract text, and use a language model to pull out structured facts. Every fact must carry a quote and a source link. Code checks that the quote exists, that the source and accession match, that the schema and values are valid, and that the language is factual. Facts that pass are published; facts that fail the hard checks are dropped; facts with weak evidence are marked Unverified. GDLG does not review individual facts before publication. Customer analysts review results in their own Workstation.

1. Sources

SEC EDGAR. Submissions metadata and recent filing records for the AI universe, plus primary documents for 8-K, 10-K, 10-Q, 6-K, DEF 14A, and 13G/13G-A. We access EDGAR under the SEC's fair-access policy: identified requests and no more than 8 requests per second from our own ceiling, below the SEC's published maximum.

Other public sources. GLEIF legal-entity identifiers (parent relationships are pending); OFAC sanctions lists; CFTC Commitments of Traders; Federal Reserve H.15 rates; BLS labor and price observations; Treasury Fiscal Data; Federal Register documents; and OpenFIGI instrument identifiers. BEA is not used because it requires a UserID.

What we do not use.

• Paywalled or login-only content.
• Social media.
• Non-public information.
• Market price data.
• The St. Louis Fed's FRED service, which bars commercial use and storage.
• Datasets whose terms do not allow our use.

Timing. The raw collector runs every six hours. The public panel is rebuilt on a weekly cadence in this stage. A customer that wants daily or on-demand freshness runs the same pipeline on its own machine.

2. The AI universe

Companies are included if they file with the SEC and their filings describe a business in at least one of ten layers: chips; foundry and equipment; memory; servers; networking; data centers; power; neoclouds; hyperscalers; AI software. Layer assignment uses SIC codes, segment disclosures, and business descriptions from the universe file. A company can belong to more than one layer.

Inclusion is not a judgement about a company. A hardware vendor can be in the universe on the same criteria as every other company. See the hardware disclosure.

3. Documents and extraction

We store the exact bytes of each public filing in a content-addressed archive and extract normalized text from HTML documents. A language model reads the text and returns structured facts: events, agreements, financing, customers and suppliers, risk factors, governance, ownership, and related disclosures.

The model is an extraction method, never a source. It is not asked to decide whether a company is good or bad, to score it, or to recommend a security. It is asked to report what the document says, with a quote.

4. Automated evidence gate

Every fact must pass these hard checks:

1. The quote must exist in the cited document after normalization.
2. The source URL and accession must match the document record.
3. The fact must match its JSON schema.
4. Dates and amounts must parse and must appear in the quote or structured source.
5. The statement must contain no banned evaluative language and no investment advice.

A fact that fails a hard check is dropped. A fact that passes the hard checks but has weak evidence, such as low model confidence or a single source, is published as Unverified with a visible label and source link. It is not presented as a confirmed finding.

GDLG does not run an editorial review queue for individual facts. The public site is generated automatically. Customer analysts review results in the Workstation.

5. Company pages and facts

A company page shows the company's SEC header data, recent filings, a three-year timeline, and the source-linked facts extracted from its documents. Sections include business and segments, customers and suppliers, financing and debt, leases and power, risk factors and legal proceedings, governance, ownership, and filed events.

Every fact links to the filing it came from. A section without source-linked facts is omitted or labeled pending. We do not fill an empty section with generic text.

6. Identifying who is who

We link records on exact identifiers first: SEC CIK, ticker, and FIGI where available, and LEI where an exact legal-name match exists in GLEIF. An optional local embedding tool can generate candidate matches, but it never merges them automatically. We do not merge entities without evidence. CUSIP is never stored or displayed.

Full GLEIF and OFAC resolution is pending. Where a link is unresolved, the public site leaves it unresolved rather than guessing.

7. Relationships and themes

Relationship edges between companies are pending. The public map currently shows company nodes by layer only. We will show an edge only when a filing discloses the relationship and the evidence gate accepts the supporting passage. No inferred or anonymous relationship is shown.

Theme pages and severity signals are also pending. Red flags and severity are a customer-Workstation feature, not a public-site feature.

8. Language models

The public pipeline uses an OpenAI-compatible model. The initial model is DeepSeek-V4.1-Flash; the same pipeline can run against a customer-local vLLM, SGLang, or llama.cpp endpoint. We record the model identifier, prompt version, source hash, and check results for every published fact.

The public pipeline processes public filings only. It does not receive customer watchlists, queries, notes, or holdings. In the Workstation, the customer's model runs on the customer's hardware, and the customer's analysis stays there.

9. Corrections

Send corrections to corrections@secfilingai.com. The source link on each item is the authoritative record. The public site does not promise a human review SLA for individual facts. Legal notices go to legal@secfilingai.com.

10. What we do not do

• We do not give investment advice, price targets, ratings, scores, or rankings.
• We do not publish "riskiest company" lists.
• We do not accuse anyone of fraud or wrongdoing; we report what was filed.
• We do not publish severity or red flags on the public site.
• We do not use non-public information or scrape login-protected sites.
• We do not show market prices.
• We do not collect what customers search in the private products.

11. Versions and changes

The universe, prompts, schemas, evidence checks, and this methodology are versioned. Each fact records the model and prompt version that produced it, plus the source hash. Changes are listed in the changelog with dates.