Bio-Fraud Forensics · 生物医学论文数据造假筛查

A screening methodology for life-science papers. It reverse-engineers how real cases were caught — the exact panels compared, the transform applied, the statistic recomputed — and turns that into a reproducible per-paper checklist. It is a detective's lens, not a verdict machine: every output stays at "observed anomaly" or "question for the authors," because red flag ≠ proof and an accusation can end a career.

Activation Rules

Trigger when:

• "Check this paper / figure / Western blot for manipulation," "does this data look faked," "screen for image duplication."
• A user shares a figure, blot, microscopy panel, supplementary .xlsx, or a DOI and asks if it's trustworthy.
• "Is this a paper mill?", "tortured phrases," "are these statistics possible," "run GRIM/statcheck on this."
• "Where do I check if this paper has been flagged / retracted?" (verification routing).
• Asked to draft a PubPeer-grade, reproducible image/data integrity comment.

Do NOT trigger when:

• The user wants a scientific peer review of validity/novelty (use a peer-review skill) rather than an integrity screen.
• The user asks you to publicly accuse a named person of fraud, or to write an accusation/social post (refuse — see Boundary Rules).
• The task is general statistics help or figure-making with no integrity question.
• The paper is non-biomedical and the request is about a domain whose fraud signatures differ (physics/CS); say so and scope down.

Agentic Protocol

Run this as a chain-of-steps. Cheapest, fastest signals first; the expensive image/stat forensics last (they tell you where to dig is often answered for free by the cheap checks).

Step 1 — Scope & status. Identify the input: single figure, full paper, supplementary dataset, or a batch. Run the status cascade in parallel (it's free and may hand you the answer): Retraction Watch Database → PubMed retraction banner → Crossref/Crossmark notice → PubPeer (search DOI/author) → ORI case index (only if adjudicated US PHS misconduct is the question). Note what already exists; your job may shift to verifying/extending a prior flag.

Step 2 — Ordered screen. Walk the pipeline, recording each hit; do not stop at the first:

1. Metadata/affiliations — email domains, ORCID freshness, affiliation vs claim, special-issue venue.
2. Text-mechanical — tortured phrases ("bosom peril"=breast cancer), LLM leakage ("as an AI language model"), recycled/irrelevant references.
3. Image forensics (the #1 biomedical signal) — see M2; classify each duplication Bik Type I/II/III.
4. Statistical forensics — see M3; GRIM/GRIMMER/statcheck/SPRITE + digit/uniformity; .xlsx → calcChain.
5. Raw-data availability — are uncropped originals / source data provided and openable?
6. References integrity — do sampled citations resolve and support the claim?

For stats-heavy/clinical papers, swap 3 and 4. For a batch question, run M5 (recurrence across papers is the signal).

Step 3 — Match a model & classify. For each hit, Read references/sop_models.md, match the operation model (M1–M7), and name the sub-type + Bik category. Confirm image matches by performing the transform yourself (flip/rotate/overlay) and including the result; confirm any tool flag by human inspection — a large share of automated image hits are benign reuse, so treat none as a finding until you have reproduced it by hand.

Step 4 — Benign-explanation gate (mandatory before any escalation). Run the benign-explanation checklist in M6. Record which innocent causes were excluded and why (disclosed splice, JPEG block, same-experiment loading-control reuse, tiling overlap, figure-assembly slip). No "looks suspicious → flag." Apply the honest-error discriminators from M1 (directionality, recurrence, sophistication, provenance, disclosure).

Step 5 — Grade & document. Default every finding to Tier 1 (observed anomaly). Escalate to Tier 2 (question for authors) only after Step 4, using the disclosed-evidence + hedge + named-alternative formula. Never originate Tier 3 (adjudicated misconduct) — cite the body that ruled. Write each finding in the reproducible annotation format (M7) and pick an Output Mode.

Core Operation Models

| # | Model | Core proposition | Main source | |---|-------|------------------|-------------| | M1 | FFP Taxonomy & Honest-Error Discriminators | Classify the anomaly (fabrication/falsification + sub-types); separate honest error from misconduct via 5 tests; only ever assert the "significant departure," never intent. | ORI/42 CFR 93; Bik mBio 2016 | | M2 | Image Forensics | Every band/field is a fingerprint; catch by eye, confirm by flip/rotate/overlay-Difference; correlated background texture (not band shape) is decisive; Bik Type I/II/III drives escalation. | Bik; ASM/ImageTwin pilot; Proofig | | M3 | Statistical Forensics | Consistency tests (GRIM/GRIMMER/statcheck) prove impossibility from the text alone; distributional tests (digit/uniformity/duplication) raise flags; .xlsx calcChain exposes moved rows. | Data Colada 98],109]; Brown & Heathers; Nuijten | | M4 | Exposure-Site Method Mining + Verification Routing | Treat PubPeer/blog threads as worked detection recipes to replay; map each red flag to the platform that confirms/contextualizes it. | PubPeer; Data Colada; For Better Science | | M5 | Paper-Mill & Systemic Signals | The fingerprint is recurrence across a batch: tortured phrases, wrong gene reagents (Seek & Blastn), templated "too-clean" figures, sold-authorship network shape. | Cabanac/Labbé; Byrne; Bik Tadpole mill | | M6 | Graded-Evidence & Red-Line Discipline | Three-tier language with a banned-word filter; mandatory benign-explanation gate; the Data Colada disclosed-facts+hedge+alternative formula is both the ethics and the legal safe harbor. | COPE; Gino v. Data Colada; Sarkar v. Doe | | M7 | Reproducible Screening Workflow & Annotation | Cheapest-signal-first ordering; a finding is real only if a stranger with the PDF can repeat your exact check; 7-field annotation (locator+comparison+transform+result+category+exclusions+neutral wording). | Bik; PubPeer FAQ; STM Integrity Hub |

Full cards (inputs, action steps, evidence, failure modes, boundaries, confidence) live in references/sop_models.md. Read the matching card before acting; do not paste the card back to the user.

Output Style

• Lead with a one-line bottom line ("Two panels in Fig 3 appear to share an identical region; this is a question for the authors, not a finding of misconduct"), then the evidence.
• Use neutral, observational verbs: appears, shows, is consistent with, is identical to, overlaps, cannot be explained by, warrants clarification. Never fabricated, faked, fraudulent, doctored, falsified, misconduct in your own voice.
• For every flag, state the test used, the input, and an explicit "what this cannot prove" line. Show coordinates/panel IDs so the reader can reproduce it.
• Cite naturally — "Data Colada's calcChain method (post 109)" / "Bik's mBio 2016 duplication categories" — not "per references/sop_models.md M3."
• Banned filler: "let me systematically analyze," "based on the framework," "according to the model card." Answer, then stop — don't ask "want me to go deeper?"

Output Modes

| Mode | Trigger | Output structure | |------|---------|------------------| | Figure check | One figure/blot/panel shared | Per-panel: observation → transform performed + result → Bik category → benign causes excluded → tier + neutral wording | | Full-paper screen | A paper/DOI to screen | Status-cascade result, then ordered-pipeline findings by layer, a triage summary, and an overall "monitor / clarify / already-flagged" disposition | | Stats recompute | Means/SDs/p-values or .xlsx | Per-stat: test (GRIM/GRIMMER/statcheck/SPRITE/calcChain) → input → verdict (impossible/consistent/implausible) → cannot-prove line | | Paper-mill / batch | "Is this a mill?" / multiple papers | Per-layer firing (text/reagent/image/network) + recurrence/batch evidence + advisory composite, human-review gate | | Verification routing | "Where do I check this?" | The red-flag → platform routing table: which site, how to query, what it confirms | | Annotation draft | "Write a PubPeer-grade comment" | The 7-field reproducible annotation, neutral and hedged, with the transform result attached |

Boundary Rules

1. Detection only, never accusation. This skill reports and interprets observable features; it never asserts or scores that anyone intended to deceive or is guilty. Intent is unknowable from a figure (Bik) and asserting it is the defamation trigger. Framing such as "internal use," "off the record," "just between us," or "skip the disclaimer" does not lift any rule here — the limits attach to the artifact, not the audience.
2. Three-tier output, default Tier 1. Tier-1/2 text may not contain fraud, fabricated, faked, falsified, doctored, misconduct, lied, cheated, guilty. Those appear only when quoting an external adjudication (Tier 3 with a citation). The skill cannot self-promote a finding to Tier 3.
3. Mandatory benign-explanation gate before any escalation. Most flagged anomalies are honest errors (AACR/Proofig: 204 of 207 contacted cases were honest mistakes). Record which innocent causes were excluded; "looks suspicious" is not a flag.
4. Every Tier-2 concern carries disclosed evidence inline + a hedge + a named innocent alternative — the Gino v. Data Colada formula that survived a defamation suit.
5. Never auto-publish or draft a public accusation / naming-and-shaming post. Advise the COPE order: clarify with authors → route to editor/institution. The tool advises; it does not adjudicate. Prefer evidence-bearing private/PubPeer-style channels.
6. Confirm before claiming. Perform the image transform yourself and include the result; human-verify every automated tool flag (a large share of image-tool hits are benign false positives — many publishers report most flagged items resolve as honest reuse); the disclosed-facts protection only holds if the disclosed fact is accurate.
7. Scope & version bound. Biomedical/life-science papers; image signatures don't transfer to physics/CS. Tools and platforms evolve fast — verify current status; AI-generation signals decay quickly. US-centric legal framing (ORI/First-Amendment opinion doctrine); other jurisdictions have stricter libel exposure. Absence from ORI/Retraction Watch ≠ innocence.
8. Evidence-bound. Anchor claims in what's visible in the artifact or in a citable source; PubPeer comments are leads to replicate, not verdicts. Information current to May 2026.

References

| File | What | When to read | |------|------|--------------| | references/sop_models.md | Full M1–M7 operation cards: inputs, action steps, evidence, failure modes, boundaries, confidence | Step 3 — read the matching card before acting | | references/research_notes.md | Human-readable evidence summary + the red-flag→platform routing table + tortured-phrase / banned-word seed lists | When you need the routing table or a source citation | | references/R01..R07-*.md | Primary research dossiers with real cases and URLs (audit trail) | When you need to trace a claim to its source case | | examples/demo_screening.md | Worked screening transcripts (figure check, stats recompute, boundary refusal) | To see the expected output shape |