grounded-fact-check

Verifies load-bearing claims in finished drafts against live sources and emits a correction ledger.

4|1|Updated Jul 30, 2026
One-click install
npx skills add https://github.com/radarist/structured-analytic-skills --skill grounded-fact-check-radarist
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: grounded-fact-check
Source: https://github.com/radarist/structured-analytic-skills/tree/main/skills/grounded-fact-check
Command: npx skills add https://github.com/radarist/structured-analytic-skills --skill grounded-fact-check-radarist

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Drafts often ship with fabricated-but-plausible specifics — wrong qubit counts, market figures, benchmark scores, or dates hidden behind a citation and a confident tone. This Skill re-verifies each load-bearing claim against live sources before publication, so no wrong number survives because everyone assumed someone else checked it. ## Core Features & Use Cases - Neutral-query verification: Each drafted value is looked up with a question that does not reveal the drafted value, preventing confirmation-biased searches. - Three-way claim classification: Every claim is marked confirmed, corrected, or unverifiable with a source locator, and corrections are applied to the text before publication. - Auditable fact-check ledger: Emits a per-claim ledger with mandatory totals and a zero-contradicted guarantee, so reviewers can audit every change. - Use Case: Before publishing a quantum-hardware brief citing qubit counts and a "first to ship" claim, run this check to catch that the drafted "32 logical qubits" is actually 8 topological qubits per the vendor's own announcement. ## Quick Start Fact-check this draft report before it ships — verify each load-bearing figure against live sources and give me a ledger of confirmed, corrected, and unverifiable claims.

Frequently Asked Questions about grounded-fact-check

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I fact-check a draft report before publishing?▼

Extract the 5-12 most decision-critical specifics from the draft, then look each one up with a neutral query that does not state the drafted value. Classify each claim as confirmed, corrected, or unverifiable against the sources, apply corrections, and emit a ledger with totals.

What is a neutral query in fact-checking?▼

A neutral query asks about a fact without revealing the drafted value, such as asking how many qubits a chip has rather than asking whether it has 32. This prevents the search from being anchored on a possibly wrong number and confirming it by bias.

Does a claim with a citation still need verification?▼

Yes. A citation proves a number was written down somewhere, not that it is right. Claims already carrying citations or source grades are exactly the case this check exists for, since fabricated specifics often hide behind confident sourcing.

When should I not use pre-publication fact-checking?▼

Do not use it when no draft exists and you need a fresh sourced answer, when the concern is internal arithmetic consistency rather than external accuracy, or when claims are forecasts and subjective judgements. Those cases call for answer generation, arithmetic recomputation, or provenance tagging instead.

What happens when a claim cannot be verified against any source?▼

The claim is classified as unverifiable and must be visibly hedged, tagged as an estimate, or dropped from the text. It is never left behind a citation as though confirmed, and no corrected value is invented from a vague result.