The system reasons. The clinician decides.
Veronara's Responsible AI framework is the public governance document for clinical AI in HealthOS. It defines how models are reviewed, versioned, audited, overridden, and governed by a named Clinical Advisory Board and Ethics Council.
Last reviewed:
Advisory Principle
CAB review
Model change log
Clinician override
Audit trail
Ethics Council
Training-data governance
Known failure modes
Human authorship
Regional deployment
The Advisory Principle.
Every AI output in a care-affecting context is advisory. The clinician decides. Every override is recorded with reason. The system does not act autonomously on clinical decisions.
Clinical Advisory Board review.
Every model with clinical effect is reviewed by the Clinical Advisory Board before production release. Reviews evaluate training data provenance, evaluation metrics, known limitations, failure modes, and clinical applicability. Review records are public under the model change log.
Model change log.
Every production model carries a semver, deployment date, training snapshot reference, evaluation record, and Clinical Advisory Board review link. Material changes produce a new version; silent drift is prohibited.
Clinician override.
Every AI recommendation is overridable as a single explicit action. Overrides are recorded with a reason and contribute to supervised retraining cycles. The clinician's judgment is the final authority.
Audit trail.
Every inference carries a reasoning trace. Clinicians, institutional operators, regulators, and ethics bodies can inspect what the system surfaced and why. Traces are retained per regional regulation.
Ethics Council.
A named Ethics Council provides institutional oversight. Mandate, members, and meeting cadence are documented on /trust/ethics-council.
Training data governance.
Training data provenance is documented. No training on patient data without explicit institutional authorization and purpose specification. Regional residency for training data.
Failure modes documented.
Known failure modes of clinical AI surfaces in HealthOS are published. Clinicians and institutional operators have visibility into when the system is likely to be wrong.
Human authorship of clinical content.
AI is not used to author clinical content published on Veronara surfaces without a named human author's substantive authorship. AI-assisted drafting is permitted; the named author carries responsibility for claims and framing.
Regional deployment.
AI inference and training run in-region per deployment. No cross-border transfer of clinical data for model operations without explicit institutional authorization.