Observed
A named event documented in a source: something happened, to whom, on a date, at a place.
Methodology
Evidence labels and the source hierarchy
What each evidence label means, which sources outrank which, and the rules a figure has to pass before it is published.
01
Source hierarchy
Sources are ranked by what they can establish, not by how strongly they are worded. A lower tier never overrides a higher one on the same fact.
- Tier 1 — Primary federal records
- USAspending.gov award records, enacted statutes, contract documents. Used for all dollar figures, award identifiers, and period dates.
- Tier 2 — Official estimates and agency documents
- CBO cost estimates, agency reports, inspector-general findings. Used for scored costs and stated agency intent, never as a substitute for an obligation record.
- Tier 3 — Documented NGO and medical research
- Named reports with stated methods and windows, such as the HRW/PHR custody-death analysis. Used for harm figures, always with the source's own boundary attached.
- Tier 4 — Named news reporting
- Wire and investigative reporting on documents we cannot yet read directly. Always labeled as reported, never upgraded to an established figure.
02
What each evidence label means
The label describes the evidentiary standing of a claim. It is not a severity rating, and it is never adjusted to make a record read more strongly.
- Observed
- Directly recorded in a named source.
- Attributed
- A named source states the finding.
- Allegation
- Asserted but not established.
- Disputed
- Credible sources conflict.
- Modeled
- Estimate from a stated method.
- Unknown
- Not established by sources currently indexed.
- Correction
- A corrected or superseded record.
- Relationship
- A sourced relationship between parties.
03
Evidence grades
Every receipt carries a grade. The grade describes the paperwork behind the record, not how serious the record is.
Primary official record read directly (statute, federal award record, court filing).
Named institutional source or documented investigative finding, verified against the published text.
Single credible secondary report, or a primary record with material gaps.
Claim recorded but not yet verified against a primary source. Treat as provisional.
04
Publication rules
- Any number displayed must trace to a source record with a URL and retrieval date.
- An appropriation is not an obligation; an obligation is not an outlay; a plan is not a bed.
- Sums are labeled with their coverage. A dataset total is never presented as a national total.
- Missing data reads as 'Not established' — never zero, never an estimate.
- Every receipt states both what its sources establish and what they do not.
- Correlation between spending and harm is shown as sequence, not causation.
- Corrections are versioned on the record itself, with the date they were applied.
05
Inferences we do not publish
These are prohibitions, not preferences. A record that would require one of them is published without the inference, or not published at all.
- Temporal proximity between a payment and an event is never presented as influence, causation, retaliation, or a quid pro quo.
- Conduct is not called corrupt, illegal, fraudulent, negligent, abusive, or a conflict of interest unless a cited source establishes it, in which case the finding is attributed.
- No numerical culpability, outrage, or moral score is derived from any field on this site.
- Identity is never inferred from a name alone; a match needs a corroborating identifier.
- A figure absent from a dataset is recorded as not found in sources currently indexed, never as zero.
- Automated processing may summarise, extract, or flag a candidate record, but it is never the final authority on culpability, causation, legality, or identity.
06
How the three evidence categories work together
Three categories of evidence, kept strictly apart. Every figure carries its source, its retrieval date, and what it does not prove.
Attributed
A named investigator, court, medical examiner, or strong body of evidence connects the harm to a failure. Attribution always belongs to someone, and we name them.
Modeled
A statistical estimate produced under a stated counterfactual and with stated uncertainty. A model describes a population, never a specific person.
These three categories are never blended. A modeled estimate is never presented as an observed event, and an observed event is never upgraded to an attributed finding without a named source making that attribution.
The chain, stage by stage
- Evidence
boundary
Public money: Congress appropriates detention funding — authority to spend, not money already spent.
Human cost boundary: Documented in the same detention system. Not presented as a causal attribution.
—
Rest of the method
This page is one part of the published method. The reader guide summarises all of it, and corrections are recorded on the affected record with the date applied.
Current section: Evidence labels