Claim Verification

Claim Verification

Journalism-grade fact-checking of a specific claim: decompose it, trace to origin, triangulate independently, and grade the verdict.

Why

Claims travel farther than their evidence. A repeatable verification pipeline replaces "sounds right" with a graded verdict, and does it fast enough to run routinely rather than only on the rare claim that triggers suspicion.

Use when / Don't use when

• Use when: a claim is load-bearing for a decision or about to be repeated onward; a statistic is doing heavy rhetorical work.
• Don't use when: nothing actually depends on the claim's truth.

Inputs → Outputs

• Inputs: a specific factual claim.
• Outputs: a verdict — verified, partly verified, unsupported, false, or unverifiable — with the evidence trail behind it.

Principles

• Verify the claim as stated — verifying a nearby, adjacent claim is how misinformation launders itself into credibility.
• The origin is the actual unit of verification, not the repetitions of it.
• Absence of confirmation is not refutation — "unverifiable" is an honest, legitimate verdict.
• Time-box the effort; verification depth should scale with the claim's load, not run to exhaustion on trivia.

Procedure

1. Fix the claim verbatim; run question-decomposition if it bundles several assertions — verify each separately, since the bundle's truth is their conjunction.
2. Trace to origin: follow the citation chain to the primary source. Note precisely where the chain breaks or mutates — mutation is itself a finding.
3. Audit the origin with source-credibility; if the claim is quantitative, run statistical-audit on the original number, not the retelling.
4. Run evidence-triangulation: count independent confirmations by origin-cluster, not by raw repetition count.
5. Search for the rebuttal using search-strategy's disconfirming queries: has this been debunked, corrected, or retracted?
6. Grade: verified, verified-with-distortion (true origin, mutated in transit), unsupported, false, or unverifiable. Tag residual uncertainty with epistemic-tagging.

Common mistakes

• Verifying the paraphrase instead of the claim as actually stated.
• Stopping at a "reputable" repeater instead of tracing to the actual origin.
• Treating the origin's mere existence as proof of its correctness.
• Over-verifying trivia while a genuinely load-bearing claim skates through unchecked.

Examples

• A statistic in a strategy memo before it goes to a board presentation.
• A historical "quote" attributed to a well-known figure.
• A rival's benchmark claim; a viral study result circulating without scrutiny.

Related

• [[critical-reading]] — the upstream extraction step that identifies which claims are actually worth verifying.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skyf0xx
• Source: skyf0xx/better-thinking
• License: MIT
• Homepage: https://zenbin.org/p/better-thinking

Install and usage instructions live in the source repository linked above.