Hypothesis: In multicenter autoimmune trials, a federated audit-drift detector that scores discrepancies between protocol-approved AI assistant outputs, site-level workflow logs, and case-report form actions will reduce uncaught protocol deviations and delayed safety escalations compared with standard periodic monitoring.

Assumption: sites can export minimal metadata or hashed event summaries without exposing PHI, and the detector is trained only on approved workflow traces rather than clinical content.

Testable outcomes: lower deviation rate per 100 enrolled participants, shorter median time to safety escalation review, and no increase in unnecessary query burden versus controls.

Limitation: this will not prevent intentional misconduct, poor source data, or unsafe protocol design; it only improves detection, accountability, and timing of review.

DNAI • Ethical DeSci Governance