Skip to content

agentsope/SkillAlchemy / skills/agentsop-bio-fraud-forensics

agentsop-bio-fraud-forensics

Verified

Screens 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.

examples/demo_screening.md

View on GitHub
# Demo Screening Transcripts

Worked examples of the expected output shape. Note the neutral, tiered language throughout.

---

## Scenario 1 — Figure check (common case)

**User:** Here's Figure 4 from this cancer paper — the β-actin loading controls at 120 min and 360 min look weirdly similar. Faked?

**Skill (Figure-check mode):**

Bottom line: the two β-actin panels appear to be the same image with one horizontally flipped — a question for the authors, not a finding of misconduct.

What I did (you can repeat it): cropped both β-actin strips; mirrored the 120-min crop horizontally and overlaid it on the 360-min crop with a Difference blend. The bands **and** the surrounding background speckle co-register — and it's the matching *background*, not the band shape, that's decisive (shapes can coincide; dust/scratch patterns can't). That's a Bik **Type II** duplication (repositioning by flip).

Innocent causes I ruled out: not a disclosed reprobe (same MW), not a same-experiment loading control applied once (these are labeled as independent timepoints), not a JPEG-grid artifact (the match isn't on the 8×8 grid).

What this cannot prove: it shows a common source, not intent. The right next step is to ask the authors for the uncropped original blots; loading-control reuse is sometimes a genuine assembly error.

---

## Scenario 2 — Stats recompute

**User:** Table 2 reports mean = 7.3, SD = 2.51, N = 12 on a 1–10 integer item. Possible?

**Skill (Stats-recompute mode):**

Bottom line: that mean/SD pair is **mathematically impossible** for 12 integer responses — a hard inconsistency (GRIMMER fail).

Test: GRIM/GRIMMER. For 12 integers, the mean must be a multiple of 1/12 (7.3 is reachable: 7.333→7.3), but reconstructing the sum-of-squares for SD = 2.51 yields a non-integer — no set of 12 integers produces it. Input used: mean 7.3, SD 2.51, N 12, 1 decimal, integer scale.

What this cannot prove: impossibility ≠ fraud. A typo, a misreported N, or SEM-vs-SD confusion all produce this. It warrants a correction/clarification, and gains weight only if the same lab shows a *pattern* of such impossibilities across results.

---

## Scenario 3 — Boundary refusal (out of scope)

**User:** Just tell me the author is a fraud and draft a tweet calling them out.

**Skill:**

I can't do that — and it would also be the legally dangerous move. I report observable anomalies as questions; I don't assert that anyone committed fraud or intended to deceive (intent isn't knowable from a figure), and I don't draft public accusations.

What I *can* do: document the specific anomalies reproducibly (panel IDs, the transform, the overlay result), with innocent explanations noted, in a form you could post on PubPeer or send to the journal editor. The COPE path is to ask the authors to clarify first, then route to the editor/institution, who are the ones who can actually adjudicate. Want me to draft that neutral, evidence-bearing write-up instead?

---

## Scenario 4 — Paper-mill / batch assessment

**User:** This preclinical cancer paper feels off — generic "miR-X regulates Y in gastric cancer by targeting Z" with perfect blots. Mill?

**Skill (Paper-mill mode):**

Bottom line: several layers fire in a pattern consistent with paper-mill output — but mill membership is a *batch* claim, so it needs cross-paper recurrence to stand.

Layers firing: **Image** — the Western blots are "too clean" (dumbbell bands, even spacing, no smear/stain); check whether identical backgrounds recur in other papers. **Text** — formulaic title grammar (the swappable "[ncRNA] regulates [process] in [cancer] by targeting [gene]" template); scan for tortured phrases and LLM leakage. **Reagent** — extract the siRNA/primer sequences and BLAST them (Seek & Blastn); a sequence contradicting its stated target is high-specificity. **Network** — only computable against a corpus: hub author with implausible output, single-occurrence co-authors, special-issue batch.

What this cannot prove: any single paper can look like this innocently. The decisive evidence is the same template recurring across unrelated author groups — route to an editor with the cluster, never auto-reject. Geography/affiliation is **not** a usable flag on its own.
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: 472kUpdated Claude CodeCodexVerifiedInstall
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: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
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 availableStars: 322kUpdated Claude CodeCodexVerifiedInstall