For developers and AI agents

Data sources and license

Where every fact comes from, what we never republish, the known limits, and the CC BY 4.0 terms.

Every fact in GetFunded comes from a public government record. Every published fact row carries the sha256 hash of the file it was parsed from and a locator for the exact record inside that file.

Sources

Source	Publisher	What we take	Terms
Form 990 / 990-PF e-file index and XML	Internal Revenue Service	Filings spine; 990-PF officers, grants paid, financial lines, Schedule B, Part XV application info; Form 990 Schedule I grants, core financials, Part IX split, Part VII compensation; filer-stated websites	US government work, public domain
Exempt Organizations Business Master File	Internal Revenue Service	Organization identity: name, EIN, address, subsection, foundation code, NTEE code, ruling date, asset/income/revenue amounts	Public domain
Automatic Revocation of Exemption List	Internal Revenue Service	Per EIN: revocation date, the date the IRS posted it, and the reinstatement date when there is one. Automatic revocations only (no return or notice filed for three years in a row)	Public domain
Publication 78 data	Internal Revenue Service	Per EIN: the IRS deductibility codes (the class of organization for tax-deductible gifts)	Public domain
Form ADV (IAPD)	US Securities and Exchange Commission	Adviser identity, CRD and SEC file numbers, offices, assets under management, private-fund schedules	Public domain
Form D	US Securities and Exchange Commission	Exempt-offering notices: issuer, offering amounts, related persons as filed	Public domain
SBIR/STTR awards	US Small Business Administration	Award records: agency, program, phase, amount, awardee	Public domain
ZCTA crosswalk	US Census Bureau	ZIP-to-ZCTA geography crosswalk used to normalize places	Public domain
Federal agencies and programs	This project	10 agencies and 16 non-dilutive federal programs, curated	CC BY 4.0
Entity-resolution labels	This project	Human match / not-match decisions	CC BY 4.0

URLs, cache rules and the limits of each source are in corpus/docs/DATA-SOURCES.md.

What we never republish

Contact channels appear only when the record carries publishability = 'public' and is not in the red privacy tier. Role inboxes such as grants@ can be public. A named person's address is withheld by policy.
Vendor-sourced facts are never republished. A database trigger makes it impossible to mark them public.
Facts scraped from funder websites are never republished. They may be used internally to check public records, only when a human approves.

The rule is structural: every public view joins the file's license, and a row from a non-republishable file cannot appear in a public view whatever its other columns say.

The provenance chain

code

dataset name → source URL → sha256-hashed immutable file → licence code
            → ingestion-ledger run → row (raw_file_id + source_record_locator)

raw_files has one row per distinct file, keyed by sha256, with three separate timestamps: when we fetched it, the publisher's last-modified date, and when we last parsed it. A re-parse can never make old data look fresh.
Record locators look like row:EIN=742961304 for a CSV row, or an element path inside the 990 XML for a filing value.
A published export is byte-identical on re-run, and its manifest lists the hash of every file.

Full description: corpus/docs/PROVENANCE.md.

Three doctrines the data obeys

Unknown is not closed. Application posture is open, preselected_only or unknown. unknown means the return carries no Part XV statement. No view, export or ranking may treat it as a refusal.
Missing is not $0. NULL means the line is absent from the return. 0 means the filer reported zero. The two are never coalesced.
Superseded filings are filtered. An amended return gets a new object id. Within one (EIN, return type, tax period) the greatest object id is the live filing. Losers keep their detail rows but lose their funding events, so nothing is double-counted.

Known limits

Entity resolution is not applied. Records of the same fund from different SEC sources are separate organizations. People are per-source. people and relationships stay out of the export until resolution certifies.
Not ingested: 990-EZ, 990-N, 990-T, paper returns and determination letters.
IRS standing has a date and a narrow meaning. It is read from the IRS master file, Publication 78 data and the Automatic Revocation of Exemption List, each as of the date shown with it. The revocation list holds automatic revocations only; an organization that lost its status in another way is not on it.
Grants-paid totals are 990-PF only. Giving-ranked views cover private foundations.
The IRS backlog. At any time tens of thousands of indexed returns have no published XML. They sit in the filings spine with nothing attached until a later run.
Back-year 990-PF grant rows are loaded for the newest index years only.
Form D names the issuer, never the investors. It is not a deal graph.
Inside parsed 990-PFs, Schedule B is present on about a quarter of filings, and Part XV application information is actionable on about a quarter; most of the rest state only that the foundation funds preselected organizations.
Keyword search matches names and titles only. Semantic search needs the optional embedding step.
Filers make mistakes. We publish filings as filed. Out-of-range numbers are clamped to NULL, not corrected.

The dataset is published under Creative Commons Attribution 4.0 International (CC BY 4.0). It covers our compilation: the selection, cleaning, linking, normalization, evidence ids and arrangement we add. It does not cover the underlying government records, which are public domain. You may use the raw records freely without crediting us.

When you share or reuse the dataset, credit it like this:

Data from GetFunded, Open Funder Database (https://getfunded.ai), licensed CC BY 4.0. Derived from IRS, SEC, SBA, and US Census public records.

Keep a link to the license. If you change the data, say so.

The code is Apache-2.0. The name and logo are trademarks of Morton Labs; see TRADEMARK.md.

Accuracy and corrections

We publish filings as the government published them. Filings can be late, amended or wrong. We filter superseded filings, but we do not correct the record.

If you find an error, open a data correction issue with the funder name, the EIN, the profile link, what is wrong and what you checked.

No warranty

The dataset is provided "as is", without warranty of any kind. See section 5 of the CC BY 4.0 legal code.