Fact Check

When to use

Reach for this when there's a single discrete claim on the table and you need to know if it's true — a statistic, a quote, an attribution, a date, a "study shows" assertion. Output is a verdict backed by the primary source, or an honest "unverifiable."

Not for: answering an open question (use rsr-deep-dive); rating a source overall (use rsr-source-eval); reconciling many sources into a narrative (use rsr-synthesis). This targets exactly one claim.

Method

1. Isolate the claim. Restate it as a single checkable proposition. Split compound claims — "sales doubled and profits rose" is two checks. Decision: if it's a matter of opinion or prediction, stop and label it non-factual.
2. Identify what would confirm or refute it. What primary evidence would settle it — the original study, the filing, the transcript, the dataset?
3. Trace to the primary. Follow citations upstream past the article to the original source. Decision: if you can only find restatements and never the primary, that itself is a finding (verdict: unverified).
4. Compare against the primary. Does the primary actually say this? Watch for numbers stripped of context, quotes trimmed to flip meaning, and correlation reported as cause.
5. Check currency. Was it true once but since superseded?
6. Rule. Assign True / Misleading / False / Unverifiable, quote the decisive evidence, and cite the primary.

Example

Claim: "A Harvard study found coffee cuts heart-disease risk by 50%." Isolate: single stat, attributed. Trace: the article cites a news piece citing an observational study. Primary: the study reports a ~15% association in the highest-consumption group, not causation, not 50%. Verdict: False — the 50% figure and the causal framing are unsupported; the real finding is a modest correlation. Evidence quoted from the study abstract, cited.

Pitfalls

• Stopping at a secondary source that "confirms" it without reaching the primary — you've verified an echo.
• Missing that a real number was quoted stripped of its qualifier (per-year vs. lifetime, relative vs. absolute).
• Calling a claim False when it's merely unverifiable — absence of evidence is its own verdict.
• Fact-checking an opinion or forecast as if it had a truth value.

Output format

Claim: <single proposition, verbatim>
Verdict: <True | Misleading | False | Unverifiable>

Primary source: <title / publisher> — <date> <link>
Decisive evidence: "<quote from primary>"
What the claim got wrong (if any): <e.g. dropped qualifier, correlation→cause>
As of: <date checked>