Reference Checker

This skill systematically verifies scientific literature references using web search and always outputs results in a consistent, three-tier format.

Workflow

1. Extract References

Extract and number all references to be checked from the user's input.

2. Look Up the Records

First run the lookup script on all references at once (from the skill folder; Python standard library, needs internet):

python3 scripts/lookup.py refs.bib          # BibTeX file
python3 scripts/lookup.py refs.txt          # one reference per line or paragraph
python3 scripts/lookup.py --doi 10.1038/nature14539

For every entry it prints the Crossref record behind a given DOI (with a title match score), the best Crossref and OpenAlex matches for the full reference, print vs. online year, volume, issue, pages / article number, and retractions, corrections and expressions of concern (Crossref updated-by from Retraction Watch, OpenAlex is_retracted). Read it as evidence:

• a DOI whose record has a title match below about 0.6 belongs to a different work;
• a real journal + volume + page/article number that the records assign to a different paper is the classic pattern of a fabricated reference;
• Crossref also lists later copies (posted-content, reprints, book chapters) — compare with the version of record; conference papers (e.g. NeurIPS) often have no DOI and appear in OpenAlex.

3. Web Search Where the Records Are Not Enough

Search the web for every entry without a convincing record (books, theses, reports, patents, proceedings without DOI) and to confirm anything doubtful. Search strategy:

• Primary: Title + journal + year
• Secondary: Authors + title (if no results)
• Tertiary: Search DOI directly (if provided)
• Quaternary: Patent databases / institutional repositories (for patents, reports, theses)

Verify the following:

• Authors: Full names, order, correct spelling; for "et al." check at least the first author
• Title: Exact wording
• Journal / Source: Correct name
• Year: Publication year (note online-first vs. print publication date)
• Volume: Number
• Page or article number: First page or article number (e.g., 123603)
• Document type: Article, patent, dissertation, report, conference paper, etc.
• DOI: Retrieve and validate
• URL: Retrieve direct URL to the publication

4. Output Language

Always respond in the same language as the user's request (English, German, or other).

Three-Tier Rating System

✅ — CORRECT

The reference is complete and all verified details match. Minor formatting differences (e.g., spacing, abbreviations) do not count as errors.

⚠️ — WARNING

The reference exists and is findable, but has issues that should be fixed before submission. Typical ⚠️ cases:

• Year error: Online-first year ≠ print year (e.g., online 2014, printed 2015)
• Incomplete details: Missing authors, missing volume/page numbers, missing patent numbers
• Non-standard format: Only a URL provided (e.g., institutional repository), no complete citation style
• Minor errors: Slight typos in author names, wrong initials, slightly different title
• Missing document type: Dissertation/patent/report not identified as such

🚨 — CRITICAL ERROR

Severe errors that compromise the integrity of the reference. Typical 🚨 cases:

• Not verifiable / possibly hallucinated: No match despite extensive search; plausible-sounding details with no real counterpart — often a sign of AI-generated false citations
• Wrong journal: Article exists but was published in a different journal
• Wrong authors: Completely different author group
• Wrong page numbers/article numbers: Significant deviation from actual details
• Wrong first author
• Retracted: the record has a retraction (updated-by type retraction, or is_retracted). Always 🚨, also when every detail is correct; if the text discusses the paper as a retracted study, say how to cite it (mark it as retracted, add the retraction notice).
• DOI belongs to a different work: the given DOI resolves to another paper. Give the correct DOI.

Output Format (ALWAYS follow)

Report Title

## Reference Verification Report

(English) or

## Referenzprüfungsbericht

(German)

Per Reference: Individual Rating

### [N] ✅ — CORRECT
[Brief confirmation]. DOI: ... URL: ...

### [N] ⚠️ — [ERROR DESCRIPTION IN UPPERCASE]
[Explanation of the problem and correct details].
DOI: ... URL: ...

### [N] 🚨 — [ERROR DESCRIPTION IN UPPERCASE]
[Explanation; explicitly flag hallucination suspicion if applicable].
Correct details (if a similar paper was found):
DOI: ... URL: ...

Note on hallucination suspicion (🚨): If a reference cannot be found despite multiple searches and the title, journal, volume, and pages are plausible but unverifiable, always include the warning: "This appears to be a hallucinated/AI-generated reference. Do not use without verification." (or in German: "Diese Referenz scheint halluziniert/KI-generiert zu sein. Nicht ohne Überprüfung verwenden.")

Summary Table

Always append a table at the end of the report. Columns: Number | Status emoji | Main issue.

English:

## Summary Table

| # | Status | Issue |
|---|--------|-------|
| [1] | ✅ | Correct |
| [2] | ⚠️ | [Brief description] |
| [3] | 🚨 | [Brief description] |

German:

## Zusammenfassende Tabelle

| Nr. | Status | Problem |
|-----|--------|---------|
| [1] | ✅ | Korrekt |
| [2] | ⚠️ | [Kurzbeschreibung] |
| [3] | 🚨 | [Kurzbeschreibung] |

If the input was a BibTeX file, add a section with the corrected entries ready to paste (and which keys to remove).

If multiple references have 🚨 status, append a notice block at the end of the table:

> **Key concern:** References [X], [Y], [Z] bear the hallmarks of AI-generated (hallucinated)
> citations. These should be replaced with verified sources or removed before submission.

(or German equivalent)

Complete Error Category Reference

Rules

1. Always look up — run scripts/lookup.py, then search the web where needed; never assess a reference from memory, even for well-known classics.
2. Always include DOI and URL — even for ✅ references.
3. Never skip a reference — every entry appears in the report.
4. Summary table is mandatory — it closes every report.
5. Keep language consistent — English or German depending on the request, maintained throughout.
6. For ambiguous matches try at least 3 different search strategies before assigning 🚨.
7. Distinguish page numbers vs. article numbers correctly.
8. Check online-first vs. print carefully — the correct citation year is the print publication year unless explicitly stated otherwise.
9. Patents and reports follow their own citation standards: inventor, applicant, patent number, country, date.
10. Recognize hallucination patterns: Plausible title + real journal + unverifiable pages/numbers = strong indicator of AI-generated false citation → always warn explicitly.