Skip to content
Plinth

Sources

Every source behind a figure

The spine of this site is the IRS Form 990 corpus. Not every number on it comes from a filing: some sources are official records about an organization, some are statistics about a place, and some are modeled rather than measured. Each one below gives its publisher, how it joins, and its limits.

Every row also carries its license, whether commercial use is permitted, whether attribution is a condition, and the date someone read the terms. A source whose terms do not permit the way we use it is removed from the product rather than footnoted here, which is why no row below says commercial use is not permitted.

How the filings become figures is on methodology. How well the matching works, measured, is on data quality. What we withhold is on redactions.

Filed

The organization stated this on its own IRS return.

The filer's own words on a named, linkable filing, so the only question is whether they filed accurately. Figures lag 12 to 24 months and are dated to the fiscal year they came from.

Official

Another government record about this organization, not its own filing.

Primary and authoritative, but published by someone other than the organization and often joined to it by name rather than EIN, which carries match risk. Each source has its own revision cadence, so a screen is only as current as the file behind it.

Context

A statistic about a place, joined by geography. Never about the organization.

Describes the area a recipient's address sits in, not the recipient itself. A grantee headquartered in a high-poverty ZIP may serve a wider or entirely different area, so read these as the setting a grant was made in, never as an outcome or a reach claim.

Inferred

Derived by a model from text, not read from a record.

The weakest class. A language model or an embedding produced it from prose, so it is directional: useful for finding something worth verifying, not evidence on its own. Each one keeps the source it was drawn from.

Resolved and computed

We worked out an identity the filing did not state. Arithmetic or graph analysis over the classes above. Neither appears above, because both are work done on top of the sources that do. Matching covers how identities are resolved and the ceiling on that; data quality publishes the measured precision, the reach, and the share we decline rather than guess.

A source that feeds a rendered figure belongs on this page. If one is missing, tell us at data@useplinth.com.