verify-citations

name	verify-citations
description	Use when the user wants to check whether references / citations are real — "are these citations real", "check this bibliography", "did the AI make up these references", "verify the references in this paper", "真假文献", "核对参考文献", "这些引用是不是编的". Cross-checks each reference against an authoritative source and flags entries that cannot be confirmed (a common LLM-hallucination failure mode). Does NOT fabricate confirmations.
license	MIT
metadata	{"category":"zotero-research","primary-interface":"zot MCP + academic lookup MCP"}
triggers	["citation","reference","bibliography","fake reference","hallucinated citation","参考文献","引用","真假文献","核对文献","文献核对"]

Verify Citations

The user wants to know whether a set of references is real — i.e. catch citations an LLM (or a careless author) fabricated: plausible-looking title + authors + venue that resolve to no actual paper. This is a defensive, integrity-focused task. The cardinal rule: never report a reference as verified unless an authoritative source actually returned a match. A fabricated confirmation is worse than no answer.

Where the references come from

1. "Check the references in this paper / what I'm reading" — first call zotero_get_open_reader_state (else zotero_get_selected_items) to identify the focused item, then read its reference list / full text via zot (full-text extraction). Do not guess which paper they mean.
2. A pasted list — the user pastes a bibliography or a chunk of "References". Parse it into individual entries (title, authors, year, venue, DOI if present).

Where verification happens (you do NOT have your own database)

This skill checks each reference against an external authority. Use sources in this order, falling back when one isn't wired:

1. zot (preferred — first-party data layer). If zot exposes a citation / literature lookup (search by title, DOI, or author+year, CrossRef / Semantic Scholar backed), use it. Keeps verification in the local-first data layer.
2. An academic lookup MCP, if the user has one wired into pi — e.g. Semantic Scholar / asta, PubMed, CrossRef. Search each reference by title, then by author+year, then by DOI.
3. None available — say so plainly: "No citation-lookup source is wired (zot search or an academic MCP). I can't verify these without one." Do not fall back to guessing from memory — your own recall is exactly what produces fake citations.

Per-reference verdict

For each entry, match on title (fuzzy) + first author + year + venue, and report one of:

• ✅ Verified — an authoritative source returned a clear match. Include the canonical title and a stable identifier (DOI / S2 / PubMed ID) from the source, not one you invented.
• ⚠️ Mismatch — a paper exists but a field is off (wrong year, wrong venue, author mismatch, title drift). Show the reference vs. the real record so the user can fix it. This is the classic "half-hallucinated" citation.
• ❓ Unverifiable — no match found. Say "could not confirm", not "fake": obscure, very recent, paywalled, non-indexed, or book/grey-literature items may be real yet unfound. Note the closest near-match if there is one.

Lead the report with a one-line tally (e.g. "12 refs: 9 verified, 1 mismatch, 2 unverifiable") then the per-entry detail. Keep it scannable.

Writing the result back (optional)

If the user wants the audit saved, offer to write it:

• as a child note on the focused paper → zotero_add_note_to_current_item, or
• as a standalone note (e.g. when auditing a pasted list not tied to one item) → zotero_create_standalone_note. Confirm in ~1 line; don't re-paste the whole report into chat after writing it.

Anti-patterns (do NOT do these)

1. Confirming from memory. Never mark a reference verified without a real lookup hit. If no source is wired, stop and say so.
2. Inventing identifiers. DOIs / IDs must come from the lookup result. Never synthesize a DOI to make a citation look checked.
3. Calling "not found" = "fake". Distinguish unverifiable from fabricated. Only call something likely fabricated when a thorough search across the available sources returns nothing close.
4. Reimplementing zot. Searching / full-text extraction is zot's job; this skill orchestrates it, it does not re-do it.
5. Silent truncation. If you only checked the first N of a long list (rate/time), say which were skipped — never imply full coverage you didn't do.