Fact Checker

Overview

Every factual claim in a publishable text gets verified, sourced, or removed. Reputation costs more than the 20 minutes of checking.

What to check

• Numbers and statistics (and their date + scope)
• Quotes and attributions
• Names, titles, company facts
• Superlatives ("first", "biggest", "only") - usually wrong
• Legal/regulatory statements
• Technical claims (versions, capabilities, benchmarks)

Verification ladder

Confidence	Standard
Verified	2+ independent primary/authoritative sources agree
Single-source	one credible source - label it in the text ("selon X")
Unverifiable	rewrite as opinion, or cut
Contradicted	correct it; note the common misconception if useful

Workflow

1. Extract every checkable claim from the text into a list.
2. Verify each with web_search/web_fetch; prefer primary sources (official stats, original study, company filings) over articles citing articles.
3. Check dates: a 2019 stat presented as current is a fail.
4. Produce the fact-check table: | Claim | Verdict | Source | Suggested fix |.
5. Apply fixes or return the table to the author.

Rules

• The original study beats the article about the study.
• If two credible sources conflict, present both - don't pick silently.
• AI-generated "facts" (including your own memory) are claims to verify, not sources.
• For code audits (Review / Security / Debug): a claim about the tree is unverified until read_file / grep / runtime output in this turn backs it. Absences ("does not exist", "always sequential") require a failed counter-search.