---
name: research-integrity
description: Review research provenance, attribution, authorship, consent, disclosure, and recordkeeping, and prepare transparent corrections or escalation questions from documented evidence.
---

# Research integrity

Use this Academic Research Assistant skill when a research workflow needs traceability, responsible attribution, ethical handling, or correction of the record. Apply the [research contract](../../docs/RESEARCH-CONTRACT.md). Distinguish an integrity concern, a methodological limitation, a reporting omission, and an established breach. The assistant cannot grant ethics approval or determine institutional culpability.

## Inputs and applicable requirements

Obtain the work under review, provenance records, participant or data agreements, author contributions, intended venue, and the user's concern. Identify applicable institutional, funder, disciplinary, legal, and publisher requirements from current authoritative materials when the task depends on them. Do not invent universal consent rules, authorship thresholds, or disclosure wording across incompatible settings.

Protect confidential material during the review. Use the minimum identifying information necessary and avoid copying sensitive records into general reports. Clarify whether the task is preventative review, documentation repair, or preparation of a response to an identified concern; each requires different evidence and actions.

## Review the research record

1. Trace important results from reported claims to data, materials, code, derivations, decisions, and versions. Separate raw records from cleaned data and derived outputs. Identify transformations, exclusions, manual interventions, and analysis changes that cannot be reconstructed.
2. Check attribution. Distinguish quotations, paraphrases, common knowledge, ideas, data, code, figures, and adapted materials. A citation does not itself grant permission for reuse. Preserve required license and credit information and determine the actual terms before distributing third-party content.
3. Review author and contributor statements against documented work and the applicable criteria. Do not assign authorship, remove a contributor, or assert sole ownership without the responsible people's involvement. Shared equipment or collaboration does not automatically establish the same contribution to every result.
4. Check consent, access, and reuse boundaries. Data obtained lawfully for one purpose may have restrictions on another use or on redistribution. Anonymization can reduce disclosure risk but is not a substitute for all consent or governance requirements.
5. Review transparency around preregistration, exclusions, deviations, conflicts, funding, assistance, and computational tools. Disclose consequential AI assistance according to the actual venue rules and its role; do not describe generated text or analysis as human verification.
6. Inspect presentation choices that might mislead: selective outcomes, image alteration, unexplained smoothing, omitted uncertainty, repeated data presented as independent, or a simulation described as observation. Establish what happened before inferring intent.
7. If a discrepancy is confirmed, propose a proportionate correction that preserves the original record and explains the change. Do not delete provenance, overwrite raw data, backdate a protocol, or invent an approval to make the project appear compliant.

## Research-mode boundaries

Quantitative integrity includes honest denominators, exclusions, analysis histories, and separation of confirmatory from exploratory findings. Qualitative integrity includes consent-sensitive quotation, representation of participant meaning, reflexivity, and acknowledgment of contextual limits. Do not fabricate an interview excerpt or an analytic audit trail.

Theoretical integrity includes identifying borrowed arguments, distinguishing established propositions from conjectures, and retaining counterexamples or gaps that change the conclusion. Computational integrity includes dataset and software provenance, licenses, seeds where relevant, parameter and environment records, and truthful claims about execution and reproducibility. A successful local run does not establish universal reproducibility or scientific validity.

## Deliverable and verification

Return a factual issue register with evidence location, applicable requirement, uncertainty, potential effect on the research record, and a proposed next action. Separate routine repairs from questions needing an ethics office, data steward, journal, legal adviser, or research-integrity process. Draft communications only within the requested scope; sending them requires explicit authorization.

Verify allegations against primary records and avoid claims about intent that the evidence cannot establish. Label unavailable records and unresolved requirements. Route citation mechanics to [citations and bibliographies](../citations-and-bibliographies/SKILL.md), scientific validity to [research appraisal](../research-appraisal/SKILL.md), and manuscript corrections to [academic editing](../academic-editing/SKILL.md). A completed review should show what was checked, not issue a blanket certificate of ethical compliance.