Biomedical Reference Verifier

Verify or normalize biomedical and life-science reference lists only when the task is about AI-caused reference errors: whether references really exist, whether DOI/PMID/title/authors/journal/year are wrong, whether a bibliography has false or corrupted entries, or whether references should be reformatted into APA/AMA/GB/T 7714/Vancouver without judging topical relevance. Trigger examples: 帮我核验这些参考文献是不是真实存在; 检查这批 references 里有没有 AI 编造的文献; 这些文献可能是 AI 生成的，请查一下哪些是假的; 请验证这些 APA 引用的 DOI、作者、题名是否对应; 帮我审查参考文献真实性; 检查这些 AMA references 的 DOI 和期刊年份是否正确; 请把这些参考文献改成 AMA 格式; 帮我把参考文献统一成 APA. Do not trigger for general article review, manuscript logic, literature relevance, citation support, or whether papers are suitable for the article.

Use this skill only for reference-list authenticity checking, AI-caused reference cleanup, and citation-format normalization. Do not use it to judge whether papers are relevant, suitable, or correctly used in the manuscript body unless the user explicitly asks for reference authenticity plus body cleanup.

Core Rule

Always reduce the task to this sequence:

1. Build or read biomedical-reference-verifier.records.v1.
2. Choose the lightest pipeline.
3. Run the script.
4. Report result files and retained process files.
5. Ask before expensive manual recovery or process-file deletion.

Load references/verification_policy.md before classifying authenticity errors, changing severe items, or explaining network/query policy.

Pipeline Choice

• Format only: use when the user only asks to convert or unify citation style. Run --pipeline format-only. Do not query Crossref/PubMed/OpenAlex.
• Verify: use when the user asks whether references are real, fake, wrong, AI-generated, DOI/PMID-mismatched, or metadata-corrupted. Run the default verify pipeline.
• Verify then format: use when the user asks both to check authenticity and standardize style. Run verify with the requested --citation-style; the fixed copy is generated after verification.

Do not run AI-assisted per-paper web searching by default. Stop after batch verification and title recovery, then ask the user whether to continue.

Machine Input

Prefer biomedical-reference-verifier.records.v1 JSON/JSONL. If the source is free text, convert it to records first using only values present in the user source. Do not fill missing source fields from Crossref, PubMed, OpenAlex, memory, or plausible guesses.

Minimal record:

json

{
  "schema": "biomedical-reference-verifier.records.v1",
  "records": [
    {
      "index": 1,
      "source": {
        "original_text": "exact source reference",
        "title": "title from source, or empty string",
        "authors": ["First Author"],
        "year": "2024",
        "journal": "Journal from source",
        "identifiers": {"doi": "10.xxxx/example", "pmid": "", "urls": []},
        "source_lines": [12],
        "context": ""
      }
    }
  ]
}

Use --input-mode records for standardized JSON/JSONL. Use Markdown worksheet and doi-context modes only for compatibility.

Commands

Run from this skill directory:

bash

python3 scripts/verify_references.py refs.md --output-dir /tmp/reference-audit --citation-style ama

Common variants:

bash

python3 scripts/verify_references.py refs.md --pipeline format-only --citation-style ama
python3 scripts/verify_references.py records.json --input-mode records --citation-style ama
python3 scripts/verify_references.py refs.md --pubmed-mode off --openalex-mode off
python3 scripts/verify_references.py refs.md --doi-output append
python3 scripts/verify_references.py refs.md --keep-process-json
python3 scripts/verify_references.py refs.md --cleanup-process-files all
python3 scripts/verify_references.py refs.md --cleanup-process-files normalized_input,extracted_references

Output Contract

Result files:

• reference-audit-summary.md: concise chat-ready summary.
• reference-audit-detail.md: detailed report with evidence links.
• references.auto-fixed.md or document.auto-fixed.md: fixed/formatted copy; never overwrites the source.

Process files retained by default:

• reference-normalized-records.json: machine-readable records.
• reference-normalized-input.md: human-readable parser inspection table.
• references.extracted.md: raw references extracted before verification.

Process files skipped by default:

• reference-audit.json: write only with --keep-process-json.

Never delete user-requested result files. Delete only process files generated by the script in the selected output directory, and only when the user chooses deletion or --cleanup-process-files is explicit.

Closeout

After running:

1. Paste or summarize reference-audit-summary.md.
2. Link result files and retained process files.
3. State which process files were cleaned or skipped.
4. Ask whether to delete retained process files:
   • A. delete all retained process files
   • B. keep selected process files
   • C. keep all process files for now
5. If severe items exist, ask whether to continue AI-assisted recovery, delete, keep with warning notes, or stop with the report.
6. Ask whether recovered DOI values should be appended when DOI output is optional.