run6270/dajia

耿同学学术打假与科学叙事方法论

Use This Skill For

Use this skill when the task is to:

• Investigate a paper, raw dataset, figure, author identity, academic title, funding claim, or institutional response for possible academic-integrity problems.
• Turn a complex biomedical, clinical, public-health, or health-tech claim into a public-facing Chinese explainer.
• Draft a video/article script that is sharp, understandable, and evidence-bounded rather than merely sensational.
• Prepare an official complaint, journal query, institutional memo, or self-check checklist from public evidence.

Do not use it to declare guilt from vibes, attack private people, or imitate a creator's personality. The skill distills a method: public-interest lead selection, evidence reconstruction, anomaly testing, confidence labeling, and story assembly.

For source anchors and corpus notes, consult references/evidence-base.md only when you need provenance.

Non-Negotiables

1. Read the whole available record before compressing it. Do not rely on titles, clips, thumbnails, isolated comments, or one article summary when full paper, supplementary data, raw data, official statements, or complete transcripts are available.
2. Separate fact, anomaly, inference, allegation, and unresolved question. Use the weakest accurate label.
3. Separate people and roles: first author, corresponding author, student, PI, lab head, hospital, school, journal, funder, publicity office, and commercial party may have different responsibility and evidence standards.
4. Protect low-power actors. Treat students, junior researchers, patients, and whistleblowers differently from high-status, high-benefit, high-control actors.
5. Prefer public, reproducible evidence. If the lead comes from a tip, convert it into documents, screenshots, raw data, paper passages, figure comparisons, official records, or repeatable checks.
6. Do not make medical, legal, or disciplinary conclusions beyond the evidence. When naming individuals or institutions in a public-facing accusation, recommend legal/editorial review.

Core Workflow

1. Define The Claim

Write one sentence for the claim under examination:

• What exactly is being claimed?
• Who benefits if the claim is accepted?
• Which public interest is at stake: grant money, degree/title, patient safety, student burden, clinical guidance, public science literacy, or institutional credibility?
• What would change your mind?

If the claim is vague, narrow it before investigating. A usable target is "Figure 2B and Figure 5D appear to reuse the same image region after rotation" or "the denominator implied by these percentages cannot produce the published decimals", not "this lab feels fake".

2. Build An Evidence Ledger

Create a ledger before writing the story:

Item:
Source:
Public URL or archive:
What it proves:
What it does not prove:
Confidence label: fact / anomaly / inference / allegation / unresolved
Follow-up check:

Fill it from:

• Paper text, supplementary files, raw data, trial registration, PubMed/journal pages, author contribution statements, retraction notices, funding pages, and institutional announcements.
• Media reports, interviews, company pages, product claims, social posts, and official responses.
• Tipster material only after converting it into independently checkable public or archived evidence.

3. Run Integrity Checks

Use the relevant checks; do not force every case into every category.

Data checks

• Repeated values across groups, rows, columns, experiments, samples, batches, or papers.
• Suspicious arithmetic: exact multiples, linear progressions, mirrored sequences, impossible percentage denominators, copied standard deviations, or precision beyond instrument/sample reality.
• End-digit and decimal-pattern anomalies: unnatural terminal-digit distribution, repeated small decimals, excessive identical precision, or format changes that match editing rather than measurement.
• Biological consistency: sex, species, tissue, cell line, sample size, time point, dosage, and assay conditions.
• Raw-data consistency: raw files must match figures, summary statistics, sample labels, time ordering, and methods. Raw data is not an automatic defense.

Image checks

• Reused panels, rotated/flipped/cropped regions, duplicated animals/cells/gels/blots, repeated backgrounds, and reused shapes with changed signals.
• Splicing, inconsistent boundaries, inconsistent scale bars, mismatched surrounding context, compression artifacts, and impossible overlay/layer behavior.
• Cross-paper reuse by the same group, collaborator, or template-like data provider.

Paper and identity checks

• Article type: research article, review, comment, letter, editorial, preprint, abstract, conference item, or correction.
• Authorship role: first author, corresponding author, co-first, co-corresponding, middle author, contributor, supervisor, hospital/school role.
• Title and honor claims: official academy membership, society fellow, talent-plan status, grant status, "院士" wording, foreign institution naming, and publicity exaggeration.
• Count claims: number of papers, impact, citations, "top journal", "breakthrough", "first in China/world"; verify against source pages rather than press-release wording.

Mechanism, clinical, and product checks

• Cell/animal/human/clinical layers are different evidentiary layers. Do not let animal or cell data imply patient benefit without clinical evidence.
• Distinguish mechanism plausibility, statistical significance, effect size, practical usefulness, safety, alternative treatments, and commercial claim.
• For health products and longevity/AI/brain-computer topics, separate what is already done, what is theoretically possible, what is legally/ethically allowed, and what is science fiction.

4. Classify Confidence

Use explicit labels in notes and drafts:

• Fact: directly supported by a public source.
• Anomaly: observable mismatch or pattern that needs explanation.
• Strong inference: multiple independent checks point to the same explanation, but no official finding yet.
• Allegation: a claim made by a source, tipster, or media; attribute it.
• Unresolved: evidence is incomplete, contradictory, or outside available access.

Prefer "this pattern is hard to explain as independent measurement" over "they fabricated it" unless the evidence and process support that stronger claim.

5. Choose The Action Ladder

Pick the lowest-risk path that can still produce accountability:

1. Internal reconstruction: ledger, figure map, data table, reproducibility notes.
2. Quiet verification: repeat a key experiment/check, ask domain peers, or run an independent statistical/image review.
3. Official channel: school academic committee, hospital, journal, funder, regulator, or company compliance contact.
4. Self-check window: for mass or systemic cases, give institutions a chance to review before public dumping when that improves correction and reduces numbness.
5. Public script/article: publish only what can be shown, label uncertainty, and avoid private harassment vectors.

For repeated-experiment proposals, make the experiment narrow, key-node focused, technically fair, and designed to test reliability rather than humiliate people. Account for student workload and negative-result handling.

Story Assembly

The storytelling style is not "be sarcastic first"; it is "make a technical contradiction legible, then return to evidence".

Use this structure:

1. Concrete hook: open with one contradiction, impossible number, reused image, exaggerated title, or public claim that a normal person can understand.
2. Source reset: state exactly what document, paper, figure, raw file, interview, or announcement is being examined.
3. Technical unpacking: explain the check in plain language. Use analogies only to clarify the technical point, not to replace it.
4. Boundary sentence: say what the evidence proves and what it does not prove.
5. Responsibility map: distinguish student/operator, PI/corresponding author, institution, journal, funder, publicity channel, and commercial beneficiary.
6. System layer: connect the case to incentives, evaluation pressure, supervision failure, grant/title reward, medical-commercial pressure, or public-science misunderstanding.
7. Action ask: name the concrete next step: correction, retraction review, data release, repeated experiment, audit, public explanation, policy change, or self-check.

Useful analogy patterns:

• Turn abstract numeric logic into embodied common sense, such as "a denominator cannot jump like this" or "precision cannot exceed the ruler".
• Turn institutional logic into roles and incentives, not personal morality alone.
• Use humor as a door, then move back to evidence; never let the punchline become the proof.

Output Formats

Evidence Memo

Claim:
Public-interest reason:
Sources reviewed:
Key facts:
Anomalies:
Checks performed:
Alternative explanations:
Responsibility map:
Confidence:
Recommended next step:
Risks and wording limits:

Public Script

Title:
Opening contradiction:
What source we are reading:
Step-by-step evidence:
What this proves:
What it does not prove:
Who should respond:
Systemic lesson:
Call for correction/self-check/review:

Official Complaint Or Query

Recipient:
Paper/project/person/institution:
Exact concern:
Source links:
Reproducible check:
Why it matters:
Requested action:
Attachments:
Contact and confidentiality preference:

Guardrails

• If the evidence is only a pattern, call it a pattern.
• If a result could be an honest labeling, cropping, or production error, say so and identify the check that would distinguish it.
• Do not infer misconduct from commercial income, media attention, nationality, institution type, personality, or public dislike.
• Do not reveal non-public personal information or encourage harassment.
• For patient-facing or health-product topics, include a medical-risk caveat and avoid treatment advice.
• For legal/disciplinary accusations, preserve exact wording, screenshots, archives, timestamps, and source URLs.