📝 PubMed Trivia Editorial Team

Our goal is to deliver primary scientific literature to general readers both accurately and accessibly. The editorial team uses Claude (Anthropic) as an adaptation-assistant tool while publishing articles.

Editorial Responsibility

  • Site operation and editorial responsibility: Kobayashi (Editor-in-chief)
    Responsible for editorial policy, prompt design, NG-term list maintenance, article sampling review, and error-correction response. Operates the workflow that uses AI (Claude) as an adaptation-assistant tool, enforces PMID + source URL disclosure in every article, and runs automated checks for medical-claims-law (薬機法) violations.
  • Contact: sciencepubmedjp@gmail.com
  • Response time: within 48 hours (excluding weekends/holidays)

Editorial Approach

PubMed Trivia is not simply "a site that lets AI summarize papers." Our editorial team designs, maintains, and reviews each step of the following workflow:

  1. Paper selection query design: The editorial team defines PubMed E-utilities search queries per category (psychology/biology). Case reports, retracted papers, and opinion pieces are excluded; systematic reviews, meta-analyses, and comprehensive reviews are prioritized.
  2. Adaptation prompt design: The system prompt passed to Claude is written and maintained by the editorial team. Three strict constraints are enforced: "Do not deviate from the original abstract's facts," "correlation ≠ causation," and "no promotional language that would violate medical-claims law (薬機法)."
  3. NG-term list maintenance: Terms like "guaranteed to cure," "lose weight," or "prevent disease" that would violate pharmaceutical-affairs or representation-of-prizes law are listed, and output containing them triggers regeneration.
  4. Sampling review: Regular random sampling of published articles to check for misreading of the source, grey-area medical claims, and naturalness of translation. Problems trigger immediate correction.
  5. Error-correction response: Reader reports are verified within 48 hours; articles are corrected if needed and dateModified is updated.

Claude (AI) — Role and Limitations

Article adaptation uses Anthropic's Claude Sonnet 4.5 and Haiku 4.5. AI is used strictly as "an assistant for adapting original-abstract facts into Japanese/English prose," following these principles:

  • AI output is always published with the PMID and source URL side-by-side, so readers can verify primary sources.
  • The prompt systematically suppresses the AI's tendency to upgrade "correlation" to "causation."
  • The prompt explicitly forbids using any information outside the original abstract, to prevent hallucinated facts or numbers.
  • Even so, we cannot guarantee zero AI errors, which is why we welcome reader error reports and commit to responding within 48 hours.

Ongoing Article-Quality Improvements

Published articles continue to receive editorial-quality improvements. See Editorial Policy → Update Log for details. For example, on 2026-10-11 the editorial team added four analytical sections (Background, Mechanism, Limitations, Reader implications) to 297 of 625 articles, bringing every article to over 2,000 characters.

Intended Audience

  • Undergraduate and graduate students (as an entry point for thesis-related literature search)
  • Researchers and healthcare professionals (for catching up on fields outside one's specialty)
  • Science-interested general readers (who want to verify the primary source behind "latest research" news items)

Article Review Process

Articles on this site are not "publish and forget." We run the following continuous review cycle:

  1. Sampling review (monthly): Each month, 10-20 articles are randomly selected and read carefully by the editorial team. We check for divergence from the original abstract, grey-area medical claims, imprecise correlation-vs-causation handling, and translation awkwardness. Problems are fixed on the spot, and if a pattern applies to other articles we audit them in a batch.
  2. Reader error reports (ad hoc): Emails to sciencepubmedjp@gmail.com are verified within 48 hours. If a correction is needed, we rewrite the article body, update dateModified in frontmatter, and ping the sitemap so Google re-crawls promptly.
  3. Quality uplift batches (as needed): We detect older articles with thin content (below a character threshold) and run a bulk pass where the editorial team's system prompt asks Claude Sonnet 4.5 to add four sections (Background, Mechanism, Limitations, Reader implications). Executed on 297 articles on 2026-10-11 (all articles now exceed 2,000 characters).
  4. Structured-data review (quarterly): We validate all schemas (Article / MedicalScholarlyArticle / BreadcrumbList / FAQPage / Organization / Person / WebSite+SearchAction) via schema.org Validator and Google Rich Results Test. All items confirmed valid on 2026-10-13.
  5. Automatic retraction handling: We periodically reference PubMed's RetractionWatch data; if a PMID we cover becomes retracted, we promptly withdraw the article or add a prominent "This article is based on a retracted paper" notice.

Our stance on expert review

This site is not an official medical institution's position, nor a specialty-society review. Accordingly, we do not claim "doctor-reviewed" or "society-endorsed." Instead:

  • Every article includes the source paper's PMID and DOI/URL, so readers can verify primary information directly
  • Adaptation prompts strictly enforce "do not deviate from the original abstract's facts", with prompt design plus sampling review as a double safeguard against the model fabricating new numbers or claims (hallucination)
  • For statements that touch on medical judgment, we add a "consult a physician" line so that articles are not read as medical advice

Related Pages