Method
How the facts are checked
Every fact in an investigation is tied to at least one document. This is the method as it actually runs, in the order it runs.
Three kinds of evidence
| Class | What it is | Marker | Sources in published investigations |
|---|---|---|---|
| Primary | The filing, docket or record itself, retrieved and hashed. | filled | 328 |
| Derived | A stored database query with its result rows and the filing key they trace to. | filled | 285 |
| Reported | Media or a data provider describing a primary record. It stays marked open until a primary record is found. | open | 80 |
Counted at build on Oct 26, 2026 across 11 published investigations: 112 verified facts, 112 of them backed by a primary or derived record (100%), and 693 distinct sources. A source cited by more than one investigation is counted once per investigation.
The steps
- 01A claim starts as a lead
A tip, a pattern in the data, or a sentence in a dossier is a lead. A lead is never evidence and never appears on an investigation.
- 02Patterns come from the full record
We work from the complete IRS 990, FEC and corporate-registry corpus in a private database, looking for shared addresses, officers, vendors and payers. The public site never queries that database; only conclusions move across.
- 03One claim, one sentence
Each lead is cut into atomic claims: a single testable statement such as who paid whom, how much, and when. Narrative is built from these, never the other way round.
- 04Registers are searched by exact key first
A filing number is searched before a name: EIN for the IRS and ProPublica, committee id for the FEC, docket number for court records. Names are searched second, together with every alias we hold for the entity.
- 05Who paid whom is checked everywhere
For any organization or person researched, we also check the FEC itemized disbursements and receipts. Many payees are nonprofits or law firms with no FEC committee of their own, so the payment is visible only on the payer's schedule.
- 06Sources are tiered by rule
Every source is labeled primary, derived, corroborating or unverified by what it is, not by how convincing it reads. Press coverage is corroborating even when the outlet is respected.
- 07Models propose, they do not write sources
Software may suggest claims from a document, and a claim is kept only when its quoted passage appears word for word in the source. A model never writes a URL, a title or a publisher; those come from the record.
- 08Mechanical checks come before verification
An item is not marked verified until it has a public address (or a stored file), a stated claim it supports, a filing key to trace it to, and an entity attached. Private paths and local network addresses are refused.
- 09Derived means a stored query
A derived item is a database query kept with the exact rows it returned and the filing key those rows trace to, so a reader can run the same lookup against the filing.
- 10Each fact has one identity
A finding carries a stable origin key and a normalized fingerprint. A second copy of the same fact is merged into the first, never published twice.
- 11Figures are compared to the register row
Where a finding states an amount, a date or a payee, the fact-checking job compares them to the register row itself: the amount within rounding, the date within the filing period, the payee against known aliases. A mismatch is flagged for a person; it never refutes a finding on its own.
- 12A refuted finding taints what depends on it
If a finding is shown wrong, every finding and investigation built on it is flagged and blocked from publishing until a person reviews it.
- 13The publish gate
An investigation publishes only when it links at least one finding, every linked finding is verified and untainted, and every listed source is publicly reachable.
- 14Checks run before every release
Validators scan for internal paths, drafting notes, unsourced claims and broken source records before the site is built.
- 15Every item keeps its custody trail
Each stored document records when and how it was retrieved, a SHA-256 hash of what was retrieved, and each later change. The hash prefix is printed in the Sources list.
Registers that count as primary
- IRS Form 990 filings and the GivingTuesday 990 index
- FEC itemized schedules (A receipts, B disbursements, I grants) and the FEC committee master
- CourtListener and PACER dockets and filed documents
- State corporation and registry records
- The IRS Business Master File of exempt organizations
What never counts
- Notes in our research vault and any dossier, ours or someone else's: they are leads
- Model output of any kind
- Encyclopedias and indexes (tertiary sources)
- Anything not yet reviewed (unverified tier)
Found something wrong? Tell us which number or sentence and what record it should rest on. Corrections are made in the record and the page is rebuilt; a changed finding re-runs the checks above.Back to the investigations or see who the records name.