radiology-research-integrity

Radiology Research Integrity

Use this skill to audit responsible research conduct across proposals, protocols, analyses, manuscripts, figures, presentations and published records. It owns the integrity classification and evidence-preservation packet. It does not decide whether misconduct occurred, investigate a person, contact an institution or publisher, or replace an authorized research-integrity process.

Non-negotiable classification boundary

Use exactly one of these labels for each concern:

• OBSERVATION — a directly inspected record, file property, version difference or attributed statement, described without motive or accusation;
• INTEGRITY_SIGNAL — an evidence-anchored inconsistency or pattern that needs clarification, provenance reconstruction or independent review;
• ALLEGATION — an attributed assertion that prohibited conduct occurred; preserve who made it, when and in what record, but never restate it as fact;
• INSTITUTIONAL_FINDING — a document-verified determination issued by an authorized institution, funder, regulator, court or publisher. Record its issuer, scope, date, status and exact source.

Do not upgrade one state to another from visual suspicion, model output, similarity software, anonymous commentary, author silence or persuasive language. Honest error and disagreement remain possible until the authorized process determines otherwise.

Choose the mode

Mode	Use when	Required result
triage	a possible integrity concern is raised with limited material	bounded classification, preservation actions, missing records and safe next owner
authorship-disclosure	authorship, CRediT, acknowledgements, COI, sponsor role or AI use is disputed or incomplete	versioned contribution/disclosure ledger and unresolved author decisions
provenance-audit	text, data, code, statistical output, images or figures may not trace to their source	transformation chain, discrepancies, claim impact and repair/escalation gate
publication-correction	a preprint, article, abstract, repository or deck may require correction, notice or withdrawal	publication-state map and evidence packet for the authorized publisher/institution
institutional-handoff	an integrity signal or allegation needs formal local handling	minimal factual handoff, preserved records, confidentiality boundary and no verdict

For the classification and escalation rules read references/evidence-states-and-escalation.md. For authorship, CRediT, COI, sponsor and AI-use questions read references/authorship-disclosure-and-selective-reporting.md. For data, text, code, DICOM and figure provenance or publication correction read references/provenance-and-publication-record.md. Use references/source-registry.md to verify the authority and currency of policy-facing claims; local and venue rules still require live verification.

Input passport

Record before judging:

request and mode | artifact IDs/versions/digests | study and Claim IDs | public/submitted/private state | institution/jurisdiction/funder/venue | applicable policy and access date | people and roles named by supplied records | evidence locators | confidentiality/access constraints | current safety or publication risk | user authority for any file or external action.

Unknown identities, intent, authorship eligibility, contribution, consent, approval, provenance or policy applicability remain NOT_VERIFIED. Review is read-only unless the user explicitly requests an in-scope artifact edit. Never upload confidential manuscripts or evidence to an external service without authority.

Required workflow

1. Freeze the review surface. Inventory exact versions, preserve originals and hashes where available, and separate public records from confidential material. Do not ask the user to alter or delete the source record.
2. Classify, do not accuse. Write atomic rows as OBSERVATION, INTEGRITY_SIGNAL, ALLEGATION or INSTITUTIONAL_FINDING; include source, alternative explanations and what would resolve it.
3. Audit the relevant chain. Reconcile authorship/contribution/disclosures; protocol and registration to analysis and reporting; raw data to derived values; source images/DICOM to panels; code/config/run to statistics; preprint to version of record and later notices.
4. Apply the radiology gate. Preserve original pixel/DICOM identity and document window/level, crop, orientation, laterality, resampling, intensity adjustment, annotation, compositing and AI generation/editing. A display transformation is not automatically improper; an undocumented or selectively applied transformation is an integrity signal, not a finding.
5. Assess claim and safety impact. State which result, participant protection, clinical claim, grant, publication or downstream artifact may be affected. Do not infer intent.
6. Choose repair or escalation. Ordinary provenance gaps may be repaired with records, correction and transparent disclosure. Potential fabrication, falsification, plagiarism, retaliation, active participant risk, evidence destruction or a material public-record problem triggers STOP_AND_ESCALATE to the authorized owner.
7. Prepare the bounded packet. Use templates/integrity-audit-report.md and, when relevant, templates/authorship-coi-ai-ledger.md. Preserve neutral language, access controls, provenance and unresolved alternatives.

Stop and escalation gates

Return STOP_AND_ESCALATE when evidence may be destroyed or altered; confidential records cannot be handled safely; participant, animal or public safety may be affected; a protected test or source record has been overwritten; a material publication claim is unreliable; retaliation or coercion is reported; or a formal allegation requires institutional procedure. Name the role, not an invented person: research-integrity officer, institutional counsel, IRB/ethics owner, data-protection owner, funder or publisher as applicable.

The skill may organize evidence and draft a factual handoff. It must not interview witnesses, determine intent, declare guilt/innocence, impose sanctions, promise confidentiality beyond the actual channel, or issue a correction/retraction itself.

Output contract

1. Audit scope and passport
2. Concern ledger — stable ID, four-state classification, exact evidence and alternative explanations
3. Authorship/CRediT/COI/sponsor/AI ledger when applicable
4. Text-data-code-image provenance map — source, transformation, output and Claim IDs
5. Selective-reporting and protocol-to-report comparison
6. Radiology display-integrity gate — original image/DICOM, transformations, laterality and PHI
7. Claim, participant and public-record impact
8. Repair / STOP_AND_ESCALATE decision — authorized owner and required evidence, never a verdict
9. Publication-record action options — correction, retraction or notice remain publisher decisions
10. Unresolved author/institution inputs and confidentiality limits

Handoffs and non-ownership

• Human/animal/biosafety approval, consent and authorized conduct -> radiology-ethics.
• DICOM de-identification, data stewardship, access and sharing -> radiology-data.
• Statistical validity or selective model evaluation -> radiology-stats / radiology-method-evaluation; whole-manuscript validity -> radiology-prereview.
• Citation support and retraction-status checking -> radiology-citation; upload package and forms -> radiology-submission; reviewer response -> radiology-response.
• Frozen executable packaging, prespecified tolerance and independent same-input replay -> radiology-reproducibility; this skill owns neutral classification and escalation when missing, altered or selectively reported run evidence creates an integrity signal, not the replay verdict.
• Institution, funder and publisher retain investigative, adjudicative and publication authority.