| name | citation-audit |
| description | Tool-driven citation verification — fetch real sources, match inline quotes, flag contradictions and suspicious domains |
| version | 1.0.0 |
| tags | ["verification","citations","quality","fact-checking"] |
| allowedTools | ["fetch","web-search"] |
| activateFor | ["verifier","citation-auditor","fact-checker"] |
Citation Audit Protocol
Your job is per-citation verification. For every citation in scope, call tools
to retrieve the real source and check it against what the report claims.
Per-citation checks (do all four)
1. URL accessibility
- Issue a
fetch against the URL
- Pass: HTTP 200 with non-empty content
- Fail: 4xx / 5xx / timeout / paywall block / redirect to unrelated page
2. Inline quote match (when inlineQuote provided)
- Search the fetched content for the exact quoted text
- Allow ≤ 5% char variance (whitespace / minor copy-edit drift)
- Fail if the quote is a paraphrase, fabricated, or attributed to a different speaker
3. Publish date sanity
- Not a future date
- Not older than 5 years for "current" claims (older OK for historical context)
- Falsified or missing dates → flag as suspicious
4. Domain reputation
- Block list: link farms, content mills, spam aggregators, AI-generated junk sites
- Prefer: gov / edu / peer-reviewed / official corporate sources / established journalism
- Treat blogs and Substack as third-tier — accept only if uniquely authoritative
Verdict statuses
| Status | Meaning |
|---|
verified | tool fetched real content; quote matches; date sane |
unverified-but-plausible | could not fetch (timeout / paywall) but claim is industry common |
unverified-suspicious | could not fetch + URL anomalies / domain on block list |
contradicted | fetched content disagrees with the quote / number |
Output JSON shape
{
"mode": "citation-audit",
"summary": {
"total": <int>,
"verified": <int>,
"unverified": <int>,
"contradicted": <int>
},
"verdicts": [
{
"index": <int>,
"url": "<source URL>",
"status": "verified" | "unverified-but-plausible" | "unverified-suspicious" | "contradicted",
"evidence": "<≥30 chars: actual snippet from source OR fetch failure reason>"
}
]
}
Hard rules (never violate)
- You MUST call
fetch for every citation. Marking verified without a tool call is dereliction.
evidence must be ≥ 30 characters and either:
- quote a real snippet from the fetched content, or
- state the specific failure (HTTP code, paywall, redirect target, date mismatch)
- If
inlineQuote is missing, status can still be verified based on URL accessibility + topical match
- A
contradicted verdict must quote both the report's claim AND the contradicting source text
- Never fabricate URLs, source titles, or quoted text — if you couldn't fetch, say so
What this skill is NOT
- Not for evaluating overall report quality (that's the meta-critic's job)
- Not for finding new sources (that's the researcher's job)
- Not for stylistic / structural review (that's the reviewer's job)
This skill is a per-citation verification gate — it operates on what the
report already cites, and either certifies or flags each one.
