AI Citation and Data Verification

Problem

LLM-written reports often mark missing reports, policies, news items, or data as empirical. Checking every citation by hand is slow, and relying on memory can either miss fabricated sources or wrongly delete real ones. cite-hallucination-check turns “can this citation be traced?” into a reviewable checking process.

How It Works

The skill extracts citations from docx, pdf, md, and plain text, then splits each item into six components: title, author/issuer, publisher, date, URL, and DOI/reference number. It first triages each citation as verifiable, reasoning-required, or non-checkable. For verifiable items, it checks local files with grep, authoritative or official sources, and web retrieval via WebSearch/WebFetch, then labels components as match, partial, not match, or unknown. It outputs a severity-ranked list across exist, exist-minor, fake, unknown, and unsure, with the closest source, mismatched component, and one-line reason.

Boundaries

It only checks citation authenticity; it does not proofread grammar, perform political review, collect industry data, or remove “AI smell.” Confidential client data or unpublished material should be verified locally and not sent to web search. Its output is evidence for review, not a final verdict; fake findings still require human confirmation before deleting or downgrading the claim.