Fact checking
Run a claim through verification and emit a verdict using a disciplined vocabulary, always with the evidence.
When to use
- The user (or a downstream step) expects a definite verdict on a statement.
- Preparing material where claims must be vetted.
- Sanity-checking research conclusions before delivery.
Tools used
fact_check— from astatement, optionalphrasesto guide verification.
Verdict vocabulary
- supported — independently corroborated.
- refuted — contradicted by evidence.
- unverifiable — insufficient/indistinct evidence (do not assert).
- misleading — true but framed to deceive.
Workflow
- Restate the statement as a claim;
fact_check. - Map raw output to the verdict vocabulary above.
- Attach the evidence trail (which sources support / contradict).
- Never ship an assertion labeled "unverifiable" or with weak confidence.
Input schema
{ "statement": "str", "phrases": "list[str]" }
Output schema
{ "verdict": "string", "confidence": "number", "evidence": "list[{source,url,quote}]" }
Security
Averaging sources is not verification; independence and freshness are. When in doubt, downgrade the verdict rather than inflate it.
Related skills
claim-verification, evidence-extraction, contradiction-search,
deep-research