# Trump Claim Tracker — Research Standard

This directory is a research corpus for a public claim tracker. It separates the factual record from any later editorial voice. Every published verdict should be reproducible from the linked evidence.

## Scope

The long-term unit of coverage is a **distinct, checkable proposition**, not every sentence or every repetition. A single speech may contain several propositions. Repeated versions of the same proposition belong to one canonical claim record with a separate occurrence log.

The initial research pass prioritizes consequential and recurrent claims from the 2024 campaign and the second Trump administration, while preserving historically important recurring claims. It is a seed corpus, not a representation that every Trump statement has already been catalogued.

## What qualifies as a claim

Include statements that are:

- Presented as factual and capable of being supported or contradicted by evidence.
- Material to public understanding, policy, public spending, civil rights, elections, health, safety, or government performance.
- Repeated often enough that one canonical assessment can clarify many occurrences.

Do not give a truth verdict to:

- Pure opinions or value judgments.
- Obvious rhetorical hyperbole without a measurable proposition.
- Policy preferences.
- Promises and forecasts whose deadlines have not passed. Track these as `promise` or `prediction` and label them `Prediction—not yet verifiable` until evidence makes an assessment possible.

When one quotation contains multiple factual propositions, split them. This follows the Duke Reporters’ Lab practice of tagging multiple claims separately and avoids hiding a supported statement and an unsupported statement under one verdict.[^1]

## Verdict vocabulary

Verdicts describe the evidence for a proposition, not the speaker’s intent.

| Verdict | Publication rule |
|---|---|
| `Accurate` | The proposition is supported by the best available evidence with no material omission. |
| `Mostly accurate` | The central proposition is supported, but a limited qualification or imprecision matters. |
| `Misleading` | Some literal support exists, but framing, omitted context, denominator, timeframe, or causal implication would likely produce a materially wrong impression. |
| `Unsupported` | The claimant has not supplied adequate evidence and reliable available evidence cannot establish the proposition. This is not automatically the same as disproved. |
| `False` | Reliable evidence directly contradicts the material proposition. |
| `Unverifiable` | Available evidence cannot reasonably resolve the proposition, often because the necessary data do not exist or the language is not operationally defined. |
| `Prediction—not yet verifiable` | The proposition concerns a future outcome or an unfinished promise. Store the stated deadline and schedule a review. |

Avoid intent-laden labels such as “lie” in research records unless there is separate, strong evidence that the speaker knew the claim was false at the time. PolitiFact likewise places the burden of proof on the speaker, rates statements using what was known when they were made, and excludes opinions from truth ratings.[^2]

## Source hierarchy

Use at least one source for the quotation and at least two independent sources for the assessment when feasible.

1. **Primary record of the statement:** White House video/transcript, the official *Compilation of Presidential Documents*, campaign transcript, full interview video, full social-media post, court transcript, or C-SPAN recording.
2. **Primary evidence:** statutes, court opinions, agency datasets, inspectors general, audited financial statements, official statistical series, contemporaneous government records, or original peer-reviewed research.
3. **High-quality synthesis:** Congressional Research Service, Government Accountability Office, National Academies, nonpartisan research institutions with disclosed methods, or systematic reviews.
4. **Independent reporting and established fact-checkers:** useful for finding the claim, identifying evidence, and checking the analysis; follow their citations back to primary material.
5. **Interested-party material:** campaigns, advocacy organizations, trade groups, and partisan outlets can document their own positions or supply leads, but should not be the sole basis for a contested verdict.

The official *Compilation of Presidential Documents* is the National Archives publication of material released by the White House Press Secretary and provides stable document identifiers.[^3] For structured interoperability, the eventual site should emit Schema.org `ClaimReview`; the Duke Fact-Check Insights guide identifies fields used across participating fact-check organizations.[^4]

## Research workflow

1. Capture the narrowest fair quotation and link to the full context.
2. Save the event date, venue, medium, transcript timestamp/page, and date captured.
3. Normalize it into a single proposition without strengthening or weakening the original.
4. Mark it `fact`, `causal claim`, `comparison`, `superlative`, `prediction`, `promise`, `number`, or `quotation/attribution`.
5. Define the metric before checking: population, geography, date range, nominal/real dollars, seasonality, and denominator.
6. Search first for primary data, including series revisions and later corrections.
7. Record evidence that supports the claim as well as evidence that contradicts it.
8. Draft a rationale that explains the decisive evidence and the most reasonable counterargument.
9. Assign a verdict only after the evidence memo is complete.
10. Require a second-person editorial review for consequential or contested assessments.
11. Publish the reviewer, status date, correction history, and archived copies where licensing permits.

## Repetition and counting

A canonical claim and an occurrence are different records. Do not inflate the number of distinct claims when the same assertion is repeated. Each occurrence should retain its own exact wording, date, venue, and source. Variants can share a canonical assessment only when the measurable proposition and relevant timeframe are materially the same.

For historical context, *The Washington Post* reported 30,573 false or misleading claim occurrences during Trump’s first term and about 750 claim variants repeated at least three times.[^5] That figure reflects that project’s definitions and occurrence-counting method; it should not be presented as the size of this tracker or combined with a distinct-claims count.

## Corrections and time-sensitive claims

- Every record has `as_of`, `last_reviewed_at`, and `next_review_at` fields.
- Do not silently overwrite a verdict. Append a dated revision explaining what evidence changed.
- Preserve the assessment that was justified by information available on the claim date, while separately displaying the current evidence status.
- For provisional agency data, note the release vintage and update after revision.
- For a prediction or promise, store its deadline and automatically queue it for review.

## Publication guardrails

- Display the exact words and full-context source beside the normalized proposition.
- Link evidence that cuts in both directions.
- Distinguish “no evidence found” from “evidence disproves it.”
- Describe data limitations prominently, especially for immigration status, crime, causation, fraud, and administrative savings.
- Never infer intent from inaccuracy alone.
- Publish a visible methodology, authorship, funding disclosure, corrections policy, and contact route.
- The project may disclose an editorial perspective, but evidence standards and verdict definitions must remain consistent across claims.

## Proposed record shape

```json
{
  "id": "tct-YYYY-NNNN",
  "canonical_claim": "A narrow, checkable proposition",
  "claim_type": "fact|causal|comparison|superlative|number|attribution|prediction|promise",
  "topics": ["economy"],
  "verdict": "Accurate|Mostly accurate|Misleading|Unsupported|False|Unverifiable|Prediction—not yet verifiable",
  "verdict_summary": "One-sentence reason",
  "as_of": "YYYY-MM-DD",
  "occurrences": [
    {
      "quote": "Exact wording",
      "spoken_at": "YYYY-MM-DD",
      "venue": "Event or publication",
      "source_url": "https://...",
      "locator": "page, paragraph, or timestamp"
    }
  ],
  "evidence": [
    {
      "stance": "supports|contradicts|context",
      "publisher": "Source organization",
      "title": "Source title",
      "published_at": "YYYY-MM-DD",
      "url": "https://...",
      "accessed_at": "YYYY-MM-DD",
      "note": "What this source establishes"
    }
  ],
  "limitations": "Known measurement or attribution limits",
  "review": {
    "researcher": "Name",
    "reviewer": "Name",
    "last_reviewed_at": "YYYY-MM-DD",
    "next_review_at": "YYYY-MM-DD"
  },
  "revisions": []
}
```

## Sources

[^1]: Duke Reporters’ Lab, “[LocalFact Report 2023](https://reporterslab.org/wp-content/uploads/sites/11/2023/03/From-Fact-Deserts-to-Fact-Streams.pdf),” 2023, methodology appendix.
[^2]: PolitiFact, “[The Principles of the Truth-O-Meter](https://www.politifact.com/article/2018/feb/12/principles-truth-o-meter-politifacts-methodology-i/),” February 12, 2018.
[^3]: U.S. Government Publishing Office and National Archives, “[Compilation of Presidential Documents](https://www.govinfo.gov/app/collection/CPD/).”
[^4]: Duke Reporters’ Lab, “[Guide to the Data — Fact-Check Insights](https://www.factcheckinsights.org/guide).”
[^5]: Glenn Kessler, Salvador Rizzo, and Meg Kelly, “ [How The Washington Post Fact Checker tracked Trump’s claims](https://www.washingtonpost.com/politics/how-fact-checker-tracked-trump-claims/2021/01/23/ad04b69a-5c1d-11eb-a976-bad6431e03e2_story.html),” *The Washington Post*, January 23, 2021.