SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillslight-research-ethics

light-research-ethics

Writing & research

View repository

Updated 10-16 06:23

About this skill

Research Ethics and Compliance Risk Review

Working Mode (Always Active)

Runs in the background in all research tasks, prompting or blocking when risks are found. When high risk is involved (fabrication/ghostwriting/ownership/privacy), clearly state that it cannot be done that way and provide compliant alternatives.

Risk Review Checklist

  1. Academic Misconduct: plagiarism, self-plagiarism, duplicate publication, fabrication/falsification of data, improper authorship. Compare against the FFP criteria of ORI/42 CFR Part 93—Fabrication, Falsification (falsification/selective omission), Plagiarism (appropriating without attributing the source); a finding of misconduct requires "significant departure from accepted practices + intent/knowledge/recklessness + preponderance of evidence," explicitly excluding honest error and academic disagreement, and does not assign labels.
  2. Data Fabrication: inventing experimental results, selective reporting, p-hacking, concealing unfavorable results.
  3. Improper Images: image reuse, improper splicing/alteration, reuse across papers (corresponding to the COPE image manipulation flowchart).
  4. Citation Norms: false citations, irrelevant citations, citation stuffing, omission of key work, citing low-quality/predatory sources (linked to m10).Retraction Check: batch-check reference DOIs against Crossref https://api.crossref.org/v1/works?filter=update-type:retraction or single article /works/<DOI>, and mark in red if a hit is found in update-to[].source=retraction-watch/publisher ; for retraction reasons, use record-id to look back up the Retraction Watch CSV.
  5. Authorship Norms: use the four ICMJE criteria to determine whether authorship is warranted (all must be met: substantial contribution + drafting/review + final approval + accountability for the whole; if only some are met → acknowledgment); use the 14 CRediT roles to generate a contribution statement; AI must not be listed as an author and must be disclosed according to use (writing assistance goes in acknowledgments; data/analysis/figures go in methods).AI policy takes values by venue: before submission, check db01 venues.csv for that venue risk_note 's ai_policy= substring (R4 has actually checked leading venues)—journals generally prohibit AI-generated images + text must be disclosed; conferences generally allow LLMs but authors are responsible for the full text, and unverified LLM citations will be rejected/retracted; for detailed red lines on AI-generated images, see m11 (light-figure-drawing) figure_integrity, and for AI disclosure templates see m07 (light-paper-drafting) mandatory_inclusions.m16 PPT image generation pipeline (imggen-enhanced) outputs only serve PPT/roadshows/frontend, and are strictly prohibited from entering the paper figure pipeline—three hard boundaries (data figures are not generated/paper figures prohibit generation/text is not baked into figures) and m16 light-slides/references/imggen_pipeline.md, m11 figure_integrity "Boundary with the R6 image generation pipeline" section are cross-referenced in three places.
  6. Privacy/Ethics Review: human-related or sensitive data are positioned under the three-tier IRB/Common Rule (45 CFR 46) review—Exempt/Expedited/Full Board; self-check using the eight criteria of 46.111 (minimization of risk, risk-benefit ratio, fair selection, informed consent and its documentation, data security monitoring, privacy and confidentiality, protections for vulnerable populations); informed consent must include purpose/procedures/risks/benefits/confidentiality/voluntary withdrawal/contact information. Chinese projects follow local ethics measures. PII is not echoed (linked to a01).Animal research involved: Research involving live animals must be approved by the Institutional Animal Care and Use Committee (IACUC, in accordance with GB/T 35892-2018 "Guidelines for the Welfare and Ethical Review of Laboratory Animals"); the approval number is at the same level as the human-subjects IRB — "no approval means a red flag"; conduct a 3R self-check (Replacement/Reduction/Refinement); distinguish routine production observation from experimental intervention — non-invasive observation such as sheep farm monitoring video is usually lower risk but institutional confirmation of whether exemption from review is required is still necessary (Decision Tree D). Farm animal welfare currently mainly consists of group standards/local standards (e.g., T/CAI 003-2019 Cashmere Goats); verify the current version when citing.
  7. Copyright/License Compliance: Do not improperly collect full copyrighted texts/templates; store only metadata/links (linked to CONVENTIONS §5). For code licenses use ScanCode (scancode --license --copyright --package <path> --json-pp out.json, can output SPDX/CycloneDX); for dependency vulnerabilities use snyk test; for supply chain malicious packages be alert to install scripts/abnormal outbound connections/obfuscation/name imitation (Socket behavioral dimensions). For work licensing use the CC Chooser four questions to select a license (attribution/commercial use/adaptation/share alike), and remind that CC is irrevocable and not for software.
  8. Exaggerated/Overpackaged Conclusions: Claims exceeding the evidence, misuse of SOTA/novel (linked to a08/m08).
  9. Paper Ghostwriting: Do not ghostwrite content intended to deceive; assistance should be collaboration rather than taking an exam on someone's behalf/fabrication.
  10. Patent Ownership: Determination of inventors/rights holders, employee inventions; the final text must be reviewed by an agent (linked to m15).
  11. Software Copyright Authenticity: The software must actually exist and the materials must not be fabricated (linked to m15).

Normative Basis and Tools (Actionable)

  • Publication Ethics: Use the ten dimensions of the COPE Core Practices (misconduct allegations/authorship/conflicts of interest/data reproducibility/ethical oversight/intellectual property/peer review/corrections and retractions, etc.) as review dimensions; for specific situations, consult the COPE Flowchart to give "who to contact next/how to handle it."
  • Authorship: The four ICMJE criteria determine "whether someone qualifies for authorship"; the 14 CRediT roles determine "what each person did."
  • Misconduct Criteria: ORI/42 CFR Part 93 FFP three elements; avoid treating honest error as misconduct.
  • Retraction Check: Crossref REST API + Retraction Watch CSV (see endpoint 4 in the checklist).
  • Similarity/AI Detection: Turnitin-type — text matching ≠ plagiarism; first exclude quotations/references/small matching passages, then look at "large unmarked" text. AI detectionvendor-reportedfalse positive rates: document level <1%, sentence level ~4%, and no flag is produced when the overall AI proportion of a whole piece is <20% (the figures come from Turnitin's official release notes, not independent third-party evaluation; check the latest page before citing, see references.md); therefore a single sentence highlighted in red is unreliable, and an AI score is only a "starting point for dialogue," not a basis for conviction; manual review + communication with the author are required; false positives are higher for non-native writing, and it must not be used as the sole basis for high-stakes judgments.Self-Plagiarism Check Implementation: Turnitin itself is closed-source and cannot be replicated, but local self-plagiarism/duplicate publication screening can be done — use scripts/text_overlap.py to compare the target text with one's own old paper corpus and find the longest verbatim consecutive overlap (corresponding to the red flag >40 words); limited to the given corpus, and must not externally claim a similarity percentage (see references.md "Offline Self-Plagiarism Check").
  • Code/License/Supply Chain: ScanCode (license + copyright + SBOM), Snyk (vulnerabilities), Socket (malicious package behavior analysis) complement each other in division of labor.
  • Work Licensing: Creative Commons Chooser four questions to select a CC license.
  • Human-Subjects Research: IRB/Common Rule (45 CFR 46.111/46.116) three-tier review and informed consent elements.
  • Privacy Compliance Workflow: Refer to the compliance skill's layered checklist + decision tree + deadline guardrails (GDPR 30 days/CCPA 45 days) to conduct data compliance review.
  • Medical Device Scenarios: When medical device software/hardware is involved, refer to ISO 13485 (design controls + risk management + CAPA) and Regulatory Compliance; the FDA QMSR final rule took effect on 2026-02-02, incorporating and replacing the old 21 CFR 820 with ISO 13485:2016. The full standard text must be purchased, and certification requires third-party audit.

For each tool's actual endpoints/parameters/known pitfalls, see the same directory references.md。

Actionable assets (directly applicable)

  • Review Report Template assets/ethics_review_template.md: Risk table (blocking/warning/advisory three levels | category | why it is a risk | compliant practice | whether third party is needed) + ICMJE/CRediT authorship and contribution statement + AI disclosure + conclusion. For compliance review, fill in this table directly.
  • Red Flag Self-Check Checklist assets/risk_checklist.md: 11 categories of risk as item-by-item checkboxes, each with red flag signals and a default starting rating; upon a hit, log it in the template and assign a level.
  • High-Stakes Decision Tree references/decision_trees.md: A suspected misconduct -> FFP three elements -> COPE flowchart; B IRB Exempt/Expedited/Full Board + 46.111 eight criteria; C pre-citation retraction check (Crossref, endpoint tested HTTP 200); D animal research involved -> experimental intervention vs routine observation -> IACUC approval number/3R (GB/T 35892-2018).
  • Batch Retraction Check Script scripts/check_retractions.py: Implements decision tree C as a runnable tool — pass a set of DOIs (or --file) to query Crossref one by one update-to[], marking 🛑retracted/⚠️correction or concern/✅no signal seen, producing a markdown or --json report,--selftest offline self-test. Two real DOIs have been tested with HTTP 200. Honest limitation: Crossref may not expose all retractions; for high-stakes literature, cross-check with Retraction Watch.consumed by m10：light-citation light-citation/scripts/verify_refs.py has inlined the same-source determination logic (same FLAG_TYPES, Crossref update-to[] + title RETRACTED prefix); during citation verification, each DOI automatically checks for retraction and reports high upon a hit; this script is for batch pre-screening or direct invocation when the full three-state table of correction/concern is needed. The two criteria must be consistent; changing the determination constants requires synchronization.
  • Offline Self-Plagiarism Check Script scripts/text_overlap.py: Implements the "more than 40 consecutive words verbatim identical" red flag as a runnable tool (pure stdlib, no dependencies) — pass the target text + a set of comparison files/globs (your own old papers/course report corpus); after normalization, use difflib.SequenceMatcher to find the longest verbatim consecutive overlapping segment (word count + original text + line locations in both files) and the overall overlap ratio (word overlap rate/Jaccard); exceeding the --min-run(default 40) threshold marks 🛑;--exclude-refs discard the references section to avoid bibliography false positives,--json/--selftest. Honest limitation: only compares the provided corpus, does not include Turnitin journals/web pages/student databases, is only for self-plagiarism and duplicate text screening, and must not externally claim a "plagiarism rate/similarity %".

Compliance review deliverables

Risk list (graded: blocking/warning/advisory) + explanation for each item (why it is a risk + compliant practice) + annotation of whether third-party (ethics committee/agent/legal) review is needed. Apply assets/ethics_review_template.md to produce.

Red lines (directly refuse and explain)

Fabricating data, deceptive ghostwriting, privacy-violating data collection, fabricating software copyright/patent materials, improperly obtaining full copyrighted texts. Provide compliant alternatives rather than cooperating.

Connections

Covers the full process of m01–m17 and a01–a09; closely coordinates with a08 (self-review), m10 (references), and m15 (software copyright and patents); risks are recorded in db09.