Scientific similarity checker

Scientific Similarity Checker

A multi-source literature search and academic integrity tool that analyzes a scientific article against global databases to surface similar work and flag potential misconduct.

---

Phase 1 — Extract Article Metadata

Before searching anything, extract as much metadata as possible from the input. The input may be:

• PDF: Use pdf-reading skill if available; otherwise read with bash pdftotext or similar
• Image: Read visually — extract title, authors, abstract from the image
• Text/paste: Parse directly

Extract these fields (mark each as found or inferred or missing):

title:          (full title)
authors:        (list: Last FM, Last FM, ...)
year:           (publication or submission year)
journal:        (journal/conference/preprint server name)
doi:            (if present)
pmid:           (if present)
abstract:       (full abstract text)
keywords:       (author-provided or extracted from abstract)
institution:    (author affiliations if visible)
funding:        (funding sources if visible)
corresponding:  (corresponding author email/name)

For the abstract, if not explicitly labeled, look for the structured paragraph immediately after author affiliations. Keywords may appear at the end of the abstract or in a separate section.

If the DOI is present, use it immediately to do a Crossref lookup (Phase 2a) — this often resolves all metadata at once.

---

Phase 2 — Multi-Database Search Strategy

Run these searches. The order matters: start with exact lookups, then broaden.

2a. Exact DOI / PMID Lookup (if available)

If a DOI is found:

GET https://api.crossref.org/works/{DOI}

If a PMID is found:

GET https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id={PMID}&retmode=json

These anchor the article definitively and may reveal if it is already published.

2b. PubMed Search (use PubMed MCP if connected, else REST API)

Run 3 separate queries, collect top 10 results each:

1. Title search: "{short_title}"[ti] (use first 8-10 significant words)
2. Author search: {first_author_last}[au] AND {year-2}:{year+1}[dp]
3. Topic search: {keyword1}[tiab] AND {keyword2}[tiab] AND {keyword3}[tiab]

See references/api-endpoints.md for full PubMed syntax reference.

2c. OpenAlex Search (free, no API key needed)

Run 3 queries:

1. Title similarity: GET https://api.openalex.org/works?search={url_encoded_title}&per_page=10

2. Author's works: First resolve author: GET https://api.openalex.org/authors?search={first_author_name}&per_page=5 Then: GET https://api.openalex.org/works?filter=author.id:{author_id}&per_page=20&sort=publication_year:desc

3. Journal/venue scan: GET https://api.openalex.org/works?filter=primary_location.source.display_name.search:{journal_name}&search={keyword1}+{keyword2}&per_page=10

2d. Semantic Scholar Search

1. Title search: GET https://api.semanticscholar.org/graph/v1/paper/search?query={encoded_title}&limit=10&fields=title,authors,year,abstract,externalIds,citationCount,venue

2. Author search (resolve author first): GET https://api.semanticscholar.org/graph/v1/author/search?query={author_name}&limit=5&fields=name,papers Then fetch the author's recent papers.

2e. arXiv Search (for preprints, CS, physics, math, biology)

GET https://export.arxiv.org/api/query?search_query=ti:{title_words}+AND+au:{first_author}&max_results=10

Also search by abstract keywords:

GET https://export.arxiv.org/api/query?search_query=abs:{keyword1}+AND+abs:{keyword2}&max_results=10

2f. Web Search (IEEE, Google Scholar, general)

Use web_search tool for:

• "{title}" IEEE — catches IEEE Xplore papers
• "{title}" site:scholar.google.com
• {first_author} {keyword1} {keyword2} {year} preprint
• "{title}" filetype:pdf

IEEE has no free API; web search is the reliable path. Read full result pages with web_fetch when snippets suggest strong relevance.

---

Phase 3 — Similarity Scoring

For each retrieved article, compute a similarity profile across four dimensions.

3a. Content Similarity (Abstract/Body)

Compare the input abstract against each retrieved abstract using textual overlap reasoning:

• Identify shared phrases ≥ 5 words (verbatim or near-verbatim)
• Note shared sentences or heavily paraphrased sentences
• Estimate overall semantic overlap: Low / Moderate / High / Very High

Thresholds:

• Low: topic overlap only, different methods/conclusions
• Moderate: same topic, some shared framing/language
• High: substantial paragraph-level overlap, same methodology
• Very High: near-identical abstracts — strong misconduct signal

3b. Topic Similarity

Compare keywords, MeSH terms, and subject matter:

• Fully same topic + same methodology = Highly Similar
• Same topic, different methodology = Related
• Adjacent topic, different field = Loosely Related

3c. Author Overlap

Count authors in common:

• All authors match: Identical authorship → self-plagiarism risk
• Majority match: Substantial overlap → check for salami slicing
• First/last author same: Core team reuse → note it
• No overlap: Independent (still flag if content is near-identical)

3d. Journal/Venue Similarity

Note if the same group has published in:

• The same journal within ±3 years → possible duplicate submission
• Closely related journals from the same publisher
• A preprint server version of what appears to be a published paper → check if properly cited

---

Phase 4 — Misconduct Assessment

Apply these tiered warning rules. Issue warnings prominently in the report.

🔴 RED FLAG — Likely Misconduct

Trigger when ANY of the following:

• Abstract similarity ≥ Very High AND author overlap ≥ 50% → Duplicate publication
• Title is identical or near-identical to an existing published paper with same authors → Verbatim duplicate
• Same data/results appear to be published in ≥2 journals without cross-reference → Salami slicing
• Unpublished manuscript matches a published article with same abstract → Plagiarism of own prior work

🟡 AMBER WARNING — Requires Investigation

• High content similarity (≥2 shared paragraphs or methods section) with author overlap
• Author has 3+ papers with highly similar titles/topics in same year
• Paper appears in a preprint server but introduction does not acknowledge the preprint
• Journal submission period overlaps with another identical submission (dual submission)

🟢 GREEN — Normal Academic Practice

• Self-citation of prior work with clear differentiation
• Same author team working on a research program (related but distinct studies)
• Preprint → journal pipeline (same paper, standard practice)
• Topic similarity without content similarity (normal scientific progress)

ℹ️ INFORMATIONAL — Context

• Other authors working in the same area (independent parallel research)
• Different language versions of same work (may be legitimate translation)

---

Phase 5 — Report Format

Always produce a structured report AND write it to a file (see Phase 6). Use this exact template:

# Scientific Similarity Analysis Report

## 📄 Article Analyzed
- **Title**: [extracted title]
- **Authors**: [list]
- **Year / Venue**: [year] | [journal/conference]
- **DOI**: [if found] | **PMID**: [if found]
- **Status**: [Published / Preprint / Unpublished manuscript]

---

## 🔍 Search Coverage
| Database | Queries Run | Results Found |
|---|---|---|
| PubMed | 3 | N |
| OpenAlex | 3 | N |
| Semantic Scholar | 2 | N |
| arXiv | 2 | N |
| IEEE/Web | 2 | N |
| Crossref (DOI) | 1 | N |
| **Total unique articles reviewed** | | **N** |

---

## ⚠️ Misconduct Assessment

[Insert RED FLAG / AMBER WARNING / GREEN / INFORMATIONAL badges here with explanations]

If no flags: "✅ No similarity patterns consistent with academic misconduct were detected."

---

## 📚 Similar Articles by Category

### By Content (Abstract / Methodology Overlap)
[List top 5, highest similarity first]
- **[Title]** — [Authors] — [Journal, Year] — [DOI/link]
  - Similarity: [High/Moderate/Low] | [Brief note on what overlaps]

### By Topic / Keywords
[List top 5-8 thematically related papers]

### By Authors (Other Work from Same Team)
[List recent papers by the same first/last author]

### By Journal / Venue
[List recent papers from same journal on related topics]

---

## 🧭 Research Landscape Summary
[2-3 sentences: Where does this paper sit in the literature? Is it consistent with a research 
program? Any unusual patterns in the publication record?]

---

## 🔗 Recommended Follow-Up
- [Any specific papers or authors worth deeper investigation]
- [Suggested searches the user could run manually]
- [Links to databases for manual review if needed]

---

Phase 6 — Write the Report to a File

After composing the report, always save it as a file in the user's current working directory (do not just print it to chat).

Filename

Build a slug from the first author's last name + year + 4-word title stub, lowercased, ASCII-only, spaces/punctuation replaced with -. Example: smith-2024-deep-learning-glioma-similarity.md.

If metadata is missing, fall back to similarity-report-YYYYMMDD-HHMM.md using the current timestamp.

Default output: Markdown

Write the full report template (the block above, with all sections filled in) to <slug>.md in the current directory. Use the write / file-creation tool available in the host; do not require /mnt/user-data/outputs/ — that path is specific to the Claude.ai web sandbox and will not exist for local CLI users.

Optional: HTML

If the user asks for HTML (phrases like "as html", "html report", "open in browser"), also produce <slug>.html. Use the first method that works:

1. pandoc (preferred — best rendering):

   pandoc <slug>.md -o <slug>.html --standalone --metadata title="Similarity Report"
   

2. Python markdown library (fallback):

   python3 -c "import markdown,sys; html=markdown.markdown(open('<slug>.md').read(), extensions=['tables','fenced_code']); print('<!DOCTYPE html><html><head><meta charset=\"utf-8\"><title>Similarity Report</title><style>body{font-family:system-ui,sans-serif;max-width:900px;margin:2em auto;padding:0 1em;line-height:1.5}table{border-collapse:collapse}td,th{border:1px solid #ccc;padding:6px 10px}code{background:#f4f4f4;padding:2px 4px}</style></head><body>'+html+'</body></html>')" > <slug>.html
   

   Install with pip install --break-system-packages markdown if missing.
3. Plain wrap (last resort): wrap the markdown in a <pre> block inside a minimal HTML skeleton.

After writing

In the chat reply, output:

• A one-line summary (e.g. "Saved similarity report → <slug>.md")
• The absolute path of every file written
• The top-level Misconduct Assessment verdict (🔴 / 🟡 / 🟢 / ℹ️)

Do not dump the full report contents into chat after writing the file — just point at it. The user opens the file when they want to read it.

---

Handling Edge Cases

Very common topic (thousands of papers): Prioritize recency (last 5 years) and exact phrase matches in title. Note that high volume is expected — focus on near-identical abstracts only.

No abstract visible: Work from title and keywords. State clearly that abstract-level similarity could not be assessed and recommend manual review of full text.

Preprint without authors: Search by title only. Flag that author-level analysis was not possible.

Non-English article: Detect the language, search in that language on PubMed (use [la] tag) and OpenAlex, and note that cross-language duplicate detection is limited.

Image input only (photo of paper): Extract as much text as possible visually. If abstract is partially cut off, note it and work with what's available.

User wants only research tool mode (no misconduct check): Skip Phase 4 entirely and focus on delivering a comprehensive related-work landscape for the topic.

---

Tips for High-Quality Output

• Always deduplicate results across databases before the final report (same paper may appear in PubMed, OpenAlex, and Semantic Scholar)
• When abstract similarity is ambiguous, quote the specific overlapping phrase so the user can judge
• For author disambiguation (common names like "Wang J"), use institution + year to narrow
• OpenAlex has the broadest coverage across all fields — prioritize it for topic searches
• PubMed is best for biomedicine; arXiv for STEM preprints; IEEE for engineering/CS
• Semantic Scholar is strong on citation graphs and has good CS/AI coverage

---

Reference Files

• references/api-endpoints.md — Full API endpoint reference for all databases with parameters
• references/misconduct-taxonomy.md — Detailed taxonomy of publication misconduct types, definitions, and authoritative sources (COPE, ORI, Retraction Watch)