agentsope/SkillAlchemy / skills/agentsop-bio-fraud-forensics
agentsop-bio-fraud-forensics
VerifiedScreens biomedical / life-science papers for signs of data fabrication, image manipulation, and statistical anomalies, using the detection techniques distilled from the field's canonical exposure platforms (PubPeer, Data Colada, Science Integrity Digest, For Better Science) and tools (ImageTwin/Proofig, statcheck, GRIM/GRIMMER, Problematic Paper Screener, Seek & Blastn). Use when asked to check a paper/figure for image duplication, blot splicing, impossible statistics, paper-mill or tortured-phrase signals, research integrity, or "is this data faked"; or when a user shares a figure, Western blot, supplementary dataset, or DOI and asks whether it looks manipulated. Reports observable anomalies as questions for clarification — it never accuses anyone of fraud.
| File | Kind | Size |
|---|---|---|
| README.md | markdown | 2.9 KB |
| SKILL.md | skill md | 12.5 KB |
| USAGE.md | markdown | 8.3 KB |
| examples/demo_screening.md | markdown | 4.2 KB |
| references/R01-misconduct-taxonomy.md | markdown | 19.9 KB |
| references/R02-image-forensics.md | markdown | 19.0 KB |
| references/R03-statistical-forensics.md | markdown | 19.3 KB |
| references/R04-exposure-sites-method.md | markdown | 21.0 KB |
| references/R05-evidence-red-lines.md | markdown | 21.4 KB |
| references/R06-screening-workflow.md | markdown | 21.7 KB |
| references/R07-paper-mill-signals.md | markdown | 19.5 KB |
| references/research_notes.md | markdown | 7.9 KB |
| references/sop_models.md | markdown | 17.4 KB |
references/sop_models.md
View on GitHub# Operation Models — Full Cards (M1–M7)
Read the card that matches the current hit before acting. Each card: When to use ·
Inputs · Action · Output · Evidence · Failure mode · Boundary · Confidence.
---
## M1 — FFP Taxonomy & Honest-Error Discriminators
**When to use:** Any time you have a candidate anomaly and must name what kind of problem it is, and whether it could be honest error.
**Inputs:** The observed anomaly + whatever provenance exists (raw files, methods disclosure, recurrence across the paper).
**Action:**
1. Place the anomaly on the FFP axis. *Fabrication* = no underlying measurement exists (invented numbers/images). *Falsification* = a real measurement was altered, relabeled, selectively omitted, or spliced. (Plagiarism is text/idea theft — adjacent, only central for paper mills.)
2. Pick the sub-type and recall its single most diagnostic signal **and** its most common benign mimic:
- Image simple duplication → pixel-identical region under two labels · mimic: figure-assembly slip (Bik Cat I).
- Image repositioned duplication → match survives flip/rotate/shift · mimic: re-photographed gel, disclosed loading-control reuse (Cat II).
- Image altered/beautified → cloned background, erased lanes, rectangular intensity steps · mimic: whole-image brightness/contrast (Cat III).
- Blot/gel splice → sharp seams, abrupt background change, irregular lane spacing · mimic: *disclosed* non-contiguous splicing with a marked dividing line.
- Statistical fabrication → GRIM/GRIMMER fail, digit/uniformity anomaly · mimic: typo, rounding, SEM-vs-SD confusion.
- Salami slicing → same cohort sliced across papers, identical N/recruitment text · mimic: properly cross-referenced secondary analysis.
- Citation/authorship manipulation → cartel topology, implausible single-year output · mimic: tight subfield, large legitimate consortium.
3. Apply the **5 honest-error discriminators**: (a) *directionality* — do all "errors" favor the hypothesis? (b) *recurrence* — one slip vs a pattern across figures/papers; (c) *sophistication* — Cat I leans innocent, Cat II/III need deliberate tools; (d) *provenance* — can dated raw files/instrument logs be produced? (e) *disclosure* — was the manipulation labeled?
4. Output only the **"significant departure"** (element 1 of the 42 CFR 93 three-element test). Intent (element 2) and the formal standard (element 3) are out of scope.
**Output:** Sub-type label + Bik category + which discriminators raise/lower the honest-error likelihood, phrased as observation.
**Evidence:** ORI/42 CFR 93 FFP definitions and three-element finding standard; Bik, Casadevall & Fang, mBio 2016 (20,621 papers, 3.8% with inappropriate duplication). Andrade ORI 2026 finding = textbook falsification-by-relabeling-and-splicing.
**Failure mode:** Asserting intent from the artifact; conflating QRP (salami, p-hacking) with FFP misconduct.
**Boundary:** US/PHS-centric; a screener establishes departure, never intent.
**Confidence:** High (regulatory + peer-reviewed).
---
## M2 — Image Forensics (the #1 biomedical signal)
**When to use:** Any photographic panel — Western blot, gel, microscopy, flow/FACS.
**Inputs:** Figure panels at max available resolution; ideally uncropped originals.
**Action (per panel):**
1. **Inventory** panels; mark halftone (blot/micrograph/FACS) vs line art. Start at loading controls (β-actin/GAPDH/tubulin) and FACS gates — the highest base-rate reuse sites.
2. **Uniqueness scan** — treat every band/field as a fingerprint; list any pair more alike than chance (same silhouette **and** same surrounding background speckle/scratches).
3. **Reveal pass** — contrast/brightness stretch + histogram equalization (expose erased bands, splice seams); emboss/edge filter (splice lines as ridges); clone-detection + noise analysis (Forensically) for copy-move and airbrush halos; ELA only as a secondary pointer.
4. **Confirm pass** — crop the suspect pair, apply flip/rotate/aspect transforms, overlay with a Difference blend (or blink-animate). Flat/black difference = same source. Decisive evidence is **correlated background texture**, not band shape (shapes can coincide; background dust/scratches cannot). Re-run on any author-supplied "originals" to defeat the rebuttal.
5. **AI-figure check** — read figure text for non-words ("testtomcels"), sanity-check anatomy/biology; for realistic generations escalate to a frequency-domain/trained detector.
6. **Classify** Bik Type I/II/III; Cat II/III rarely happen by accident → stronger clarification question (still a question).
**Output:** Per-pair: panels, transform applied + result image, Bik category, neutral statement.
**Evidence:** Bik mBio 2016; Science Integrity Digest blog; ASM/ImageTwin 12-month publisher pilot (mBio 2025, PMC12505991) — a substantial share of automated flags resolve as benign reuse, so human verification is mandatory (treat the exact published rates as the source of record, not any single quoted figure); Proofig/ImageTwin docs.
**Failure mode:** Treating an automated hit as a finding; flagging JPEG-grid artifacts; single-band match without background co-registration; trusting peer review / author AI-disclosure as a safeguard.
**Boundary:** Tools are halftone-only, database-limited, vendor accuracy unverified. An overlay match proves common source, not intent.
**Confidence:** High for method; medium for vendor tool numbers.
---
## M3 — Statistical Forensics
**When to use:** Reported means/SDs/p-values, integer/Likert endpoints, or a supplementary numeric dataset / `.xlsx`.
**Inputs:** Reported (mean, SD, N, decimals, scale bounds), APA test statistics (t/F/r/χ²/z + df), or the raw data file.
**Action — three tiers:**
1. **Consistency battery (text alone, no raw data, hard flags):**
- *GRIM* — for integer-item means, check `mean×N` is reachable; e.g. mean 5.27, N 43 → impossible.
- *GRIMMER* — checks SD/variance consistency for integer data (ignore test-3-only failures, known bug).
- *statcheck* — recompute p from statistic+df; a *gross/decision* inconsistency (significance flips) is the high-value flag.
- *Bounded-scale max-SD* — for scale [a,b], SD ≤ √((b−μ)(μ−a)); exceeding it is impossible.
2. **Distributional battery (needs a column/dataset, soft flags):** terminal-/second-digit χ² (missing human rounding or over-clustering); KS/Anderson-Darling vs the *plausible* real distribution (suspicious uniformity = RANDBETWEEN signature); near-duplicate row/block detection validated against a Monte-Carlo null; zero-variance/identical-block in controls; SPRITE reconstruction for bounded integers (impossible or implausible individual values, e.g. a child eating ~60 carrots); p-curve across a set (left-skew/spike under .05 = p-hacking signature, set-level).
3. **Metadata forensics (`.xlsx`/`.docx` supplement):** rename to `.zip`, unzip, read `xl/calcChain.xml` — formula calc-order vs visible row order; out-of-place rows = manual relocation (Data Colada "Clusterfake"). Note the sorting caveat. Also scan font/style runs (duplication tell) and sort-order/duplicate-ID violations.
**Output:** Per stat: test, input numbers, verdict (impossible / consistent / implausible / tampered), and a "what this cannot prove" line. Severity: hard flag = mathematical impossibility (GRIM/GRIMMER 1–2, statcheck gross, SPRITE no-solution, calcChain relocation); soft flag = probabilistic. Require ≥1 hard flag OR a corroborated pattern of soft flags across multiple results before escalating to a Tier-2 question for the authors (with the hedge + alternative formula) — never to a verdict.
**Evidence:** Data Colada [98] (odometer: KS-uniformity + rounding asymmetry + font-twin Monte Carlo) and [109] (calcChain); Brown & Heathers (GRIM/GRIMMER/SPRITE); Nuijten & Epskamp (statcheck — ~½ psych articles ≥1 inconsistency, ~1/8 gross). Use existing tools: R `scrutiny`, `statcheck`/statcheck.io, `rsprite2`, Python `pysprite`, `benford.analysis`.
**Failure mode:** First-digit Benford on bounded lab data; digit tests on assigned numbers (IDs, doses); GRIM on non-integer/undisclosed-item means; treating one inconsistency as proof (typos are common).
**Boundary:** Consistency tests prove inconsistency, never intent. Metadata needs the original file format.
**Confidence:** High for consistency tests; medium for distributional.
---
## M4 — Exposure-Site Method Mining + Verification Routing
**When to use:** You need a detection recipe, or you need to confirm/contextualize a red flag against the right platform.
**Inputs:** A red flag (image dup / stat anomaly / mill suspicion / status question) and a DOI/author.
**Action:**
1. **Mine, don't just look up.** Read PubPeer threads and Data Colada / For Better Science / Science Integrity Digest posts as *worked recipes* — extract the step (which panels, which transform, which recomputed statistic), then replay it on your target.
2. **Route the flag to the platform that answers it** using the routing table in `research_notes.md`. Two traditions: image-forensics (Bik, Smut Clyde, Cheshire, Tiger BB8 → PubPeer + For Better Science) vs statistical/data forensics (Data Colada, Nick Brown → Data Colada + journals).
3. **Status cascade for "already flagged/retracted?":** Retraction Watch DB (Author/DOI) → PubMed banner → Crossref/Crossmark notice → ORI (adjudicated US PHS only) → scite (citation-level). Note that status feeds lag the actual retraction.
4. Keep METHOD (replay the catch) ahead of STATUS (look up the verdict); most papers you screen are not yet retracted.
**Output:** The replayed finding + the platform(s) used + how to query them.
**Evidence:** PubPeer (Actinopolyspora biskrensis 2010 Cancer blot thread → OP-OVERLAY-BLINK); Data Colada [109]; For Better Science Smut Clyde/Tiger BB8 mill interview; Crossref-hosted Retraction Watch DB.
**Failure mode:** Treating a pseudonymous PubPeer comment as a verdict (it's a hypothesis to replicate); "just check if it's retracted" as the whole method.
**Boundary:** Coverage is uneven (skewed to high-profile, English, image+psych-stats). This card extracts method and routes; it does not adjudicate.
**Confidence:** High.
---
## M5 — Paper-Mill & Systemic Signals
**When to use:** Suspicion that a paper is one of a batch, not a one-off; or a "is this a mill?" question.
**Inputs:** The paper (text, figures, reagents) and, where possible, a comparison corpus / co-author graph.
**Action — four independent layers; agreement across layers raises confidence:**
1. **Text** — tortured phrases (synonym-swapped nonsense: "bosom peril"=breast cancer, "counterfeit consciousness"=AI, "mean square blunder"=MSE) via a dictionary scan; LLM leakage ("as an AI language model", "regenerate response"); recycled/irrelevant references.
2. **Reagent** — extract nucleotide sequences (siRNA/shRNA/primers) and BLAST them (Seek & Blastn logic); a sequence that contradicts its stated target/role is a high-specificity flag (~3.8–4% wrong in top cancer journals).
3. **Image** — "too clean" templated blots (dumbbell/"tadpole" bands, regular spacing, no smear/stain), recycled identical backgrounds across unrelated papers, "Death Star" cytometry, standardized bar-graph styling; formulaic title grammar ("[ncRNA] regulates [process] in [cancer] by targeting [gene]").
4. **Network/batch** — hub authors with implausible single-year output and short careers; ego-network clustering coefficient far below norm; co-author pairs that co-occur exactly once; citation cartels; clusters submitted to one special issue in a short window.
The decisive move is **recurrence across nominally unrelated papers** — match the target against a cluster, not in isolation.
**Output:** Which layers fired + the recurrence/batch evidence + an advisory composite that routes to human review (never auto-reject).
**Evidence:** Cabanac/Labbé Problematic Paper Screener; Byrne & Labbé Seek & Blastn (PLOS ONE 2019); Bik Tadpole mill (400+ papers); Scientific Reports 2024 authorship-for-sale network fingerprint; Hindawi/Wiley mass retractions.
**Failure mode:** Generic AI-text classifier or geography/affiliation as a standalone flag (penalizes non-native-English and early-career authors); single-paper image forensics claimed as mill membership; identity checks (buyers are real people).
**Boundary:** Network detection needs a bibliometric corpus, not a single PDF. Prevalence figures are detector-dependent estimates. AI-generation signals decay fast.
**Confidence:** High for text/reagent; medium for network (corpus-dependent).
---
## M6 — Graded-Evidence & Red-Line Discipline (ethical core)
**When to use:** Before emitting *any* finding. This is the gate that keeps the skill from becoming a defamation engine.
**Inputs:** A candidate finding from M1–M5.
**Action:**
1. **Run the benign-explanation checklist; exclude innocent causes before escalating:**
- Seam in gel/blot → *disclosed lane splicing* (check methods/legend for a marked dividing line).
- Blocky edges/halos → *JPEG/compression artifacts* (follow the 8×8 grid, appear globally; don't flag).
- Reused loading-control band → *same-experiment reuse* (acceptable; only cross-*unrelated*-experiment reuse concerns).
- Same image in two papers → *disclosed self-reuse / methods figure* (check citation/attribution).
- Overlapping microscopy fields → *tiling/stage-movement overlap* (adjacent-field edge overlap is plausible).
- Same panel twice → *figure-assembly slip* (dominant honest-error cause; Tier-2 question, never Tier-3).
- "Too clean" / transcription mismatch → *honest typo, rounding, journal template*.
2. **Assign a tier:**
- **Tier 1 — Observed anomaly** (default; the only tier the tool may originate): publicly verifiable from the paper, nothing about intent. "Fig 3B and Fig 5A appear to share an identical region (boxed)." Verbs: *appears / shows / is identical to / overlaps*.
- **Tier 2 — Question for authors** (only after the checklist fails to explain it): the Data Colada formula — (a) disclosed evidence/coordinates, (b) hedged interpretation ("suggests... may have been... we believe"), (c) a named innocent alternative. Phrased as an open question.
- **Tier 3 — Adjudicated misconduct** (NEVER originated): cite the journal retraction notice / institutional or ORI finding / court ruling. Only here may *misconduct/fabrication* appear, as a quote.
3. **Enforce the banned-word filter** on Tier-1/2 output: fraud, fraudulent, fabricated, faked, falsified, doctored, misconduct, lied, cheated, guilty.
4. **Recommend the COPE order**, do not act it out: clarify with authors → route to editor/institution. Surface the base rate ("most flagged anomalies are honest errors") on every report.
**Output:** The tiered, neutral, hedged finding with innocent causes recorded.
**Evidence:** PubPeer norms; Bik ("intent is beyond image analysis"); Gino v. Data Colada (dismissed Sept 2024 — disclosed-facts+hedge+alternative = protected opinion); Sarkar v. Doe (Michigan COA 2016 — opinion on disclosed facts not defamatory; *form* matters); AACR/Proofig base rate (204/207 honest).
**Failure mode:** "Guilt probability" score; auto-generated public accusation; naming the "most likely manipulating author"; relying on a single detector's manipulation call.
**Boundary:** The skill reaches Tier 2 at most on its own authority. US opinion-doctrine protection is weaker abroad. Protection requires the disclosed fact to be accurate.
**Confidence:** High (legal + community-norm grounded).
---
## M7 — Reproducible Screening Workflow & Annotation
**When to use:** Structuring a full-paper screen and writing findings up.
**Inputs:** The paper + the findings from M1–M6.
**Action:**
1. **Order by cheapest/fastest signal first** (the Agentic Protocol pipeline). A strong early hit raises the prior and tells you where to dig. Swap image↔stats for stats-heavy/clinical papers. Run PubPeer/Retraction-Watch lookup early in parallel.
2. **Per-step stop rules:** DISCARD a candidate if an innocent cause fully explains it; CONTINUE recording even after one solid finding; never convert a tool flag to a stated finding without human confirmation against the original.
3. **Write each finding in the 7-field reproducible annotation format:**
1. *Locator* — DOI + exact figure/panel/sub-region ("Fig 3A pT231Tau blot, lane 4" or "x≈[120–260], y≈[40–110] px at published resolution").
2. *Comparison object* — what vs what ("Fig 1B lane 2 vs Fig 5C lane 5").
3. *Transformation applied* — the operation the reader repeats ("horizontal flip + overlay").
4. *Result* — attach the overlaid/aligned image; state co-registration.
5. *Bik category* — I / II / III.
6. *Innocent-explanation check* — which were ruled out and why.
7. *Neutral description* — observation, not intent.
**Output:** A finding any stranger with the PDF can reproduce.
**Evidence:** Bik scan order ("don't read text," ~1–15 min/paper); PubPeer FAQ ("perform the transformation yourself and include the result"; compression artifacts rejected); Cassava/Hoau-Yan Wang blots (lane-count vs conditions, splice positions, boxed/floating bands, background reuse); STM Integrity Hub multi-signal pipeline.
**Failure mode:** Pixel-coordinate precision theater (annotate by lane/band/region "at published resolution"); exonerating on clean metadata (elite labs fabricate too — Masliah); accusatory wording (gets PubPeer comments rejected and isn't reproducible).
**Boundary:** A reproducible annotation is the unit of a defensible finding; without the transform+result it's a guess.
**Confidence:** High.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |