MLOps Controls for Biomedical Foundation Models
Holding
Biomedical foundation models require MLOps controls that are stricter than ordinary model registries. The controlled object includes the base model, adaptation data, prompt or task wrapper, evaluation set, deployment context, and prohibited use.
Authority
NIST AI risk guidance, FDA SaMD concepts, and high-risk AI obligations all emphasize risk, transparency, performance, and monitoring. Foundation models do not escape these duties because they are general-purpose at the pretraining stage.
Issue
The failure mode is uncontrolled reuse. A model trained for biomedical language or imaging can be repurposed for diagnosis, trial screening, triage, or research summarization without a new intended-use analysis. Capability does not equal authorization.
Resolution
The platform should maintain model cards, dataset restrictions, task approvals, evaluation gates, prompt/version control, deployment constraints, monitoring, and prohibited-use enforcement. Fine-tuning and retrieval sources should be governed with the same seriousness as model weights.
Evidence Package
The record should include pretraining provenance where available, adaptation lineage, evaluation reports, intended-use statements, risk assessment, release notes, prompt or wrapper versions, monitoring logs, and incident records. The system should prove not only what the model can do, but what it is allowed to do.