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/R02-image-forensics.md
View on GitHubStatus: pass
## Key Findings
- Image duplication/manipulation is the #1 reproducible fraud signal in life-science papers. Bik, Casadevall & Fang (mBio 2016) screened 20,621 papers and found **3.8% (782 papers)** contained inappropriate image duplication — and ~half showed features suggesting deliberate manipulation rather than honest error. This is the empirical anchor of the whole field.
- Bik's three-category taxonomy is the canonical screening rubric and maps directly to escalating intent:
- **Category I — Simple Duplication** (29.4% of flagged): identical panel reused to represent different conditions. Often honest error (e.g., wrong file pasted) but still a record problem.
- **Category II — Duplication with Repositioning** (45.5%, the most common): overlapping image that has been shifted, rotated, or flipped/mirrored relative to its twin. Less likely to be honest error.
- **Category III — Duplication with Alteration** (25.1%): partial duplication of bands/lanes/cell clusters *within or between* panels, often with rotation/reversal, stamping, patching, or FACS-gate manipulation. Most likely deliberate.
- The core human technique is **comparative visual inspection** built on a single premise: every Western blot band, gel lane, microscopy field, and FACS cloud should be *unique like a fingerprint*. Bands resembling each other more than chance allows (same "pancake/dumbbell/bowtie" shape, same background spots/scratches/airbubbles) signal duplication.
- Manual reveal techniques amplify what the eye misses: contrast/brightness stretching and **histogram equalization** expose splice seams and erased bands; **emboss/edge filters** make splice lines pop as 3D ridges; **overlay + flip/rotate** comparison confirms repositioned duplicates; **clone detection, ELA, and noise analysis** (Forensically) localize copy-paste and airbrushing.
- Automated tools are now in production at major publishers: **ImageTwin** (ASM since 2023, ~120M-figure database) and **Proofig** (Science since early 2025; also cited by Nature). Both are *triage* tools — every hit requires human verification because false positives are real and frequent.
- The 2024-2025 frontier is **generative-AI figures**. The Frontiers "rat with giant genitals" Midjourney case (Feb 2024) was the wake-up call; both ImageTwin and Proofig shipped dedicated AI-generated-image detectors in 2024-2025 that work in the **frequency domain** (GANs/diffusion leave spectral fingerprints invisible to the eye). Proofig reports ~98% detection / 0.02% false-positive for AI microscopy; these numbers are vendor-stated and not independently verified.
## Detection Cases (>=2 required)
### Case 1: Flipped β-actin loading control reused across timepoints (Category II, Western blot)
- 情境 (Situation): A Western blot figure presents β-actin loading controls for a cytosol sample at multiple timepoints (e.g., 120 min vs 360 min) as if each were an independent experiment. Loading controls are the highest-yield place to look because labs are tempted to reuse the "same actin" rather than re-run it.
- 约束 (Constraints): The two band sets are not pixel-identical at a glance — one has been horizontally flipped, defeating naive eyeballing and any duplicate-finder that only checks for exact copies. Published JPEG is low resolution; no raw files available.
- 检测步骤 (Detection Steps):
1. Treat each band as a unique fingerprint; scan the panel for any two bands sharing identical silhouette + identical surrounding background speckle.
2. Crop the two candidate band regions.
3. Apply the flip/rotate transform set: mirror one crop horizontally, then overlay on the other (Photoshop "Difference" blend or simple alpha overlay). A near-flat/black difference image = same source.
4. Confirm the giveaway is the *background noise around the band* matching, not just band shape (shape alone can coincide; correlated background dust/scratches cannot).
5. Classify as Category II (repositioning by horizontal flip).
- 结果 (Outcome): β-actin bands for 120 min were identical-but-horizontally-flipped to those at 360 min — impossible for genuinely independent runs. Reported on PubPeer as duplication with repositioning.
- 可提取的操作 (Extractable Operation): For any loading-control or repeated-control panel, run the **flip/rotate-then-overlay-Difference** check; correlated *background* texture (not band shape) is the decisive evidence.
### Case 2: AI-generated (Midjourney) anatomical figures in a published review (generative-AI fraud)
- 情境 (Situation): Guo et al., "Cellular functions of spermatogonial stem cells in relation to JAK/STAT signaling pathway," *Frontiers in Cell and Developmental Biology* (Feb 2024, DOI 10.3389/fcell.2023.1339390). Figures were declared as Midjourney-generated but passed peer review and were published.
- 约束 (Constraints): No forensic software needed — but the deeper constraint is that *more realistic* AI figures will not be this obvious. Two reviewers and an editor missed it under accelerated review. Generative figures have no "original" to duplicate-match against, so duplication detectors are blind to them.
- 检测步骤 (Detection Steps):
1. Read the text inside the figure. AI giveaway #1 is gibberish labels: "testtomcels," "senctolic," "diʒlocttal stem ells," "signal bıidimg the recetein," "proprounization." Real figures never contain non-words.
2. Check anatomical/biological plausibility: a rat drawn with four enormous testicles and an impossible penis; a "pathway diagram" with invented arrows and terms; a panel that looks like "pizzas with pink salami and blue tomatoes."
3. Cross-check the disclosure statement (here authors admitted Midjourney use).
4. For non-obvious cases, escalate to **frequency-domain AI-detection** (FFT/spectral analysis or a trained detector like Proofig/ImageTwin AI module) which catches GAN/diffusion spectral fingerprints invisible to the eye.
- 结果 (Outcome): Mocked widely on X/Twitter within hours; Frontiers retracted the paper. Became the canonical example that peer review does not catch generative fraud and that publishers needed automated AI-image screening.
- 可提取的操作 (Extractable Operation): A two-tier AI-figure check — (a) human "read-the-labels + anatomy-sanity" pass that catches crude generations instantly, then (b) frequency-domain/spectral or trained-detector pass for realistic generations. Always inspect figure text for non-words.
### Case 3 (supporting): ASM/ImageTwin production pilot — automated flag then 3-stage human verification
- 情境 (Situation): ASM Journals integrated ImageTwin and ran a 12-month pilot (Mar 2023-Mar 2024) on 2,627 accepted manuscripts to quantify how automated screening + human review performs at scale.
- 约束 (Constraints): ImageTwin only processes *halftone* images (Western blots, microscopy, photographs) — it cannot read line art, sketches, plots, or graphs, leaving coverage gaps. It flags pixel-level similarity, which produces benign hits (legitimately reused placeholders/controls).
- 检测步骤 (Detection Steps):
1. ImageTwin auto-flags pixel-level similarities within the manuscript and against its ~120M-figure published database.
2. Stage 2: a human image specialist visually inspects every flag.
3. Stage 3: confirm with Photoshop "Difference" function (overlay subtraction) before any concern is raised with authors.
- 结果 (Outcome): 410/2,627 manuscripts (15.6%) flagged; only 248 (60% of flags, 3.9% of total) were genuine duplications — most of those unintentional (reused controls/placeholders), ~5.3% reused from prior publications. The ~40% of flags that were *not* genuine duplications is the empirical false-positive load.
- 可提取的操作 (Extractable Operation): Never report a tool hit as fraud. Mandatory pipeline = **automated flag → human visual inspection → Difference-overlay confirmation**; expect ~40% of raw flags to be discarded.
## Evidence Sources (source_id, URL, type, confidence)
- S1 | https://journals.asm.org/doi/full/10.1128/mbio.00809-16 (and PMC mirror PMC4941872) | Bik, Casadevall, Fang 2016 mBio — foundational duplication-prevalence paper, Category I/II/III definitions | High
- S2 | https://scienceintegritydigest.com/2020/02/21/western-blots/ | Bik blog — "every band is unique" visual-inspection method | High
- S3 | https://scienceintegritydigest.com/2019/11/23/scanning-for-duplications/ | Bik blog — Category I/II/III restated with intent ladder, flipped/shifted/rotated examples | High
- S4 | https://scienceintegritydigest.com/2024/02/15/the-rat-with-the-big-balls-and-enormous-penis-how-frontiers-published-a-paper-with-botched-ai-generated-images/ | Bik blog — Midjourney/Frontiers AI-figure case | High
- S5 | https://intercom.help/proofig/en/articles/9414427-analyzing-suspected-issues-in-western-blot-images-with-proofig-best-practices-and-tips | Proofig docs — emboss filter, histogram equalization for blots | High
- S6 | https://intercom.help/proofig/en/articles/10544884-histogram-equalization-in-image-forensics | Proofig docs — histogram equalization caveat ("visualization tool, not proof") | High
- S7 | https://29a.ch/photo-forensics/ + https://29a.ch/2015/08/21/noise-analysis-for-image-forensics | Forensically (Jonas Wagner) — ELA, clone detection, noise analysis, level sweep, magnifier | High
- S8 | https://pmc.ncbi.nlm.nih.gov/articles/PMC12505991/ (mBio 2025, DOI 10.1128/mbio.01990-25) | ASM/ImageTwin pilot — scale results, 3-stage verification, halftone-only limitation | High
- S9 | https://www.science.org/doi/10.1126/science.adn7530 | Science 2024 editorial "Genuine images in 2024" — Proofig adoption (403 on fetch; corroborated via search) | Medium
- S10 | https://tnqtech.com/imagetwin-launches-ai-generated-image-detection-capability/ | ImageTwin AI-generated-image detection launch (frequency-domain) | Medium
- S11 | https://www.proofig.com/newsroom/ (Nature AI image-integrity feature) | Proofig AI detection ~98%/0.02% FP — vendor-stated | Low-Medium (vendor claim)
- S12 | https://arxiv.org/html/2408.13786v1 | "Localization of Synthetic Manipulations in Western Blot Images" — patch-based synthetic-vs-real heatmap | Medium
- S13 | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11323046/ | ELA-CNN integration framework for authenticity verification | Medium
- S14 | https://retractionwatch.com/2026/04/19/elisabeth-bik-scientific-sleuth-image-duplication-mbio-biorxiv-preprint/ | Retraction Watch 10-yr retrospective on Bik preprint | Medium
## Supported Candidate Operations
1. **Band/field uniqueness scan** — treat every blot band, gel lane, microscopy field, and FACS cloud as a fingerprint; flag any pair more similar than chance (matching silhouette AND matching background speckle/scratches). [S2, S3]
2. **Flip/rotate-then-overlay-Difference** — for any suspected repositioned duplicate, mirror/rotate one crop and subtract (Photoshop "Difference" or alpha overlay); a flat/black result = same source. [S1, S8]
3. **Contrast/brightness stretch + histogram equalization** — push pixel intensities to full range to expose erased bands and splice seams. [S5, S6]
4. **Emboss / edge filter** — render edges as 3D ridges to make splice lines and patched rectangles visible. [S5]
5. **Clone detection (Forensically)** — auto-locate copy-pasted regions within a single panel (Category III). [S7]
6. **Noise analysis (Forensically)** — reverse a median filter to show only noise; warping/airbrushing/transformed clones leave a tell-tale black halo / disrupted noise field. [S7]
7. **Error Level Analysis (ELA)** — recompress and highlight regions with anomalous compression error; pasted/edited regions differ from surroundings. [S7, S13]
8. **AI-figure two-tier check** — (a) read figure text for non-words + anatomy/biology sanity; (b) frequency-domain/spectral or trained detector for realistic generations. [S4, S10]
9. **Mandatory tool→human→Difference verification pipeline** — never escalate a raw automated flag without human inspection and Difference-overlay confirmation. [S8]
10. **Loading-control-first targeting** — start screening at β-actin/GAPDH/tubulin and FACS controls; these are the highest-base-rate sites of reuse. [S1, S3]
## Rejected or Weak Candidate Operations
- **ELA as standalone proof** — heavily JPEG-recompressed publisher figures and uniform backgrounds make ELA noisy and ambiguous on scientific images; useful as a pointer, never as a verdict. [S7, S13]
- **Exact-pixel duplicate matching alone** — misses Category II (flipped/rotated/shifted) and Category III (partial) duplications, which are the majority. Must be paired with transform-invariant comparison. [S1]
- **Histogram equalization as proof of fraud** — Proofig explicitly states it is a *visualization* aid, not evidence; equalization artifacts can mimic manipulation. [S6]
- **Vendor accuracy numbers as ground truth** — Proofig's 98%/0.02%-FP AI-detection figures are self-reported and not independently benchmarked; treat as marketing. [S11]
- **Trusting peer review / disclosure** — the Frontiers case shows reviewers and editors pass obvious AI fraud; author disclosure of AI tools is not a safeguard. [S4]
- **Metadata / EXIF reliance** — publisher pipelines strip metadata; absence of edit history proves nothing for figures.
## Domain-specific Patterns
- **Western blots / gels**: splice lines (sharp vertical edges, abrupt background-intensity change between lanes), erased bands (suspiciously clean rectangles), duplicated/flipped bands, "too perfect" beautified bands lacking natural background noise. Loading controls (β-actin, GAPDH, tubulin) are the #1 reuse site.
- **Microscopy / histology**: duplicated cell clusters within/between fields (Category III), overlapping fields presented as different conditions (Category II), cloned regions to hide artifacts.
- **Flow cytometry / FACS**: identical scatter-plot regions or gates relabeled with different percentages (a Category III signature called out in Bik 2016).
- **Intent ladder**: Category I → possible honest error; Category II/III → repositioning/alteration is hard to do by accident, raising probability of intent. This ladder should drive how strongly a screener phrases a concern.
- **AI-generated figures**: gibberish text/labels, anatomically impossible structures, invented pathway terms, surreal color/texture; realistic ones detectable only via frequency-domain spectral fingerprints.
## Boundaries and Uncertainties (esp. tool false positives)
- **Tool false positives are large and structural.** ASM pilot: ~40% of ImageTwin flags were NOT genuine duplications (410 flagged, 248 real). Legitimately reused placeholders and shared controls trip the matcher. Mandatory human verification. [S8]
- **Coverage gaps.** ImageTwin processes only halftone images (blots, microscopy, photos) — it ignores line art, plots, schematics, graphs, where data fabrication can also hide. [S8]
- **Database dependence for cross-paper duplication.** ImageTwin (~120M figures) and Proofig can only flag reuse from *indexed* sources; reuse from unindexed/non-English/paywalled or pre-publication figures is invisible.
- **Proofig vs ImageTwin trade-off.** Proofig is reported to give fewer false positives on Western blots and is more accurate there, but is slower and historically had a smaller cross-source database than ImageTwin. Different publishers chose differently (ASM→ImageTwin, Science→Proofig). [search corroboration; S8, S9]
- **AI-detection is an arms race.** Adversarial perturbations in the frequency domain can defeat AI-image detectors; vendor accuracy figures are unverified and likely degrade on out-of-distribution / post-processed images. [S10, S11, arXiv 2407.20836]
- **Human inspection is subjective.** Bik uses three independent reviewers reaching consensus; a lone screener risks both false positives (coincidental band shape) and misses. Background-texture correlation, not shape, is the more reliable discriminator.
- **Resolution ceiling.** Published low-res JPEGs lose the high-frequency detail many techniques (ELA, noise analysis) rely on; raw/TIFF source is far stronger evidence when obtainable.
- **An overlay match proves common source, not intent.** A duplication finding establishes that two regions share a source; whether it is fraud vs honest file-handling error is a separate, evidence-weighted judgment.
## Recommendations for Later Skill Compilation
1. **Encode Bik's I/II/III taxonomy as the core classifier** with the intent ladder (I = possible error → II/III = likely deliberate), and make the screener output the category explicitly.
2. **Make "every band/field is a fingerprint" the guiding heuristic**, and instruct that *correlated background texture* (speckle, scratches, dust) — not band/cell shape alone — is the decisive evidence of duplication.
3. **Codify the per-figure step list** (see below) as a deterministic checklist, panel by panel.
4. **Hard-wire the verification pipeline**: any automated or eyeball flag → human inspection → flip/rotate-overlay-Difference confirmation before escalation. Build in the expectation that ~40% of raw hits are false positives.
5. **Add a dedicated AI-figure module**: tier-1 read-the-labels + anatomy-sanity; tier-2 frequency-domain/trained detector. Flag any non-word text in figures.
6. **Ethics/calibration layer**: outputs must say "possible duplication / common source," never "fraud"; distinguish honest-error patterns (reused control, wrong-file paste = Category I) from deliberate patterns; require multiple independent looks for high-stakes claims.
7. **Document tool boundaries** so the skill never over-trusts software: halftone-only coverage, database limits, vendor-claimed accuracy, AI-detection arms race.
### Step-by-step procedure a screener runs on ONE figure panel
1. **Inventory** every panel and sub-panel; note which are halftone (blot/micrograph/FACS) vs line art.
2. **Catalog uniqueness**: for each blot band / cell field / FACS cloud, note shape + surrounding background texture. List any pair that looks more alike than chance.
3. **Contrast/brightness stretch + histogram equalization** on each halftone panel → look for splice seams, abrupt background steps between lanes, suspiciously clean rectangles (erased bands), "too perfect" beautified bands.
4. **Emboss/edge filter** → splice lines and patched rectangles appear as ridges.
5. For each suspect pair, **crop and overlay**; apply flip + rotate transforms; use Difference blend. Flat/black = same source (record as Cat I if as-is, Cat II if repositioned).
6. **Clone detection + noise analysis (Forensically)** on single panels → localize copy-paste within a panel (Cat III) and airbrush/warp halos.
7. **ELA** as a secondary pointer to recompressed/pasted regions (never as sole proof).
8. **AI-figure check**: read all figure text for non-words; sanity-check anatomy/biology; if realistic and still suspect, run frequency-domain/trained AI detector.
9. **Optional automated pass** (ImageTwin/Proofig) for within-paper and cross-database duplication; then human-verify every hit.
10. **Classify and phrase**: assign Bik category, state "possible common source / duplication," note whether pattern fits honest error vs deliberate, and record exact pixel coordinates + transform that produced the match for reproducibility.
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 |