Research integrity audit

Research Integrity Audit

Evaluate evidence, not people. Preserve a chain of custody for every source artifact.

Workflow

1. State the scope: paper version, supplement, figures, data, code, external records, and unavailable artifacts.
2. Create a finding record before assigning severity. Include exact location, source values or image panels, method, output, assumptions, alternatives, and impact.
3. Reproduce deterministic observations with a second method when possible.
4. Separate these evidence levels and assign a compatible verification status:
   • verified reporting or numerical error;
   • identification or method problem;
   • logic or argument problem;
   • reproducibility gap;
   • unresolved anomaly requiring source data or clarification;
   • official finding documented by an identified authority.
5. Search for extraction, typesetting, rounding, versioning, or honest implementation explanations.
6. Assess whether multiple findings are independent or manifestations of one underlying issue.
7. Validate the report before sharing it:

python -m paper_audit validate-report audit-report.json

Screening routes

• For figures, read references/image-and-data-screening.md.
• For quantitative plots whose geometry carries numerical evidence, use $figure-visual-consistency-audit; return here to classify the combined record.
• For evidence language and escalation thresholds, read references/evidence-taxonomy.md.
• For a direct empirical social-science request, start with $social-science-paper-audit rather than treating image-oriented screening as sufficient. If that orchestrator invoked this skill, classify the evidence and return control; do not invoke the orchestrator again.

Rules

• Never convert an automated alert directly into a misconduct finding.
• Never use a single probability score for “fraud.”
• Do not use “fabricated,” “falsified,” “fraudulent,” “manipulated results,” deliberate intent, or “misconduct” as a conclusion unless a formal record establishes it; otherwise describe the document-level contradiction.
• Attribute an official action precisely with authority, ISO date, HTTP(S) url, and a bounded scope. Keep the ordinary finding claim neutral; official misconduct language belongs only in official_record.scope.
• Use verified for verified_reporting_error and official_finding; use unresolved or blocked for unresolved_anomaly. Follow the complete compatibility table in references/evidence-taxonomy.md for the other classes.
• “No issue detected” means only that the performed checks did not find one.
• Keep humor, insults, and speculation out of reports that may affect identifiable researchers.

Validator boundary

The JSON Schema and deterministic validator close the report model, reject prohibited score fields recursively, bound text and collection sizes, enforce classification/status combinations, and screen several high-risk allegation patterns. They are not a complete natural-language review. Regexes can miss paraphrases or mishandle context, and the validator does not fetch a URL or prove that prose matches an official record. Before publication, a human must read the entire report, open every official source, verify its authority and date, and confirm that official_record.scope does not broaden the source.

Output

Return a scope statement, artifact inventory, findings grouped by evidence class, checks that passed, unresolved questions, central-conclusion impact, and proportionate next actions.