Model versioning is critical for reproducibility, rollback capability, and compliance. MLflow Model Registry concepts: registered models (named entities with versions), versions (specific model states), stages (Staging, Production, Archived). Workflow: 1) Train model, log to MLflow with metrics and artifacts. 2) Register promising model version: mlflow.register_model(model_uri, "MyModel"). 3) Add model card annotations: description, intended use, training data lineage, performance metrics by slice. 4) Transition to Staging for testing: client.transition_model_version_stage(name, version, "Staging"). 5) Run automated evaluation pipeline on hold-out test set. 6) Human approval (ML engineer sign-off). 7) Transition to Production: replace previous version. 8) Monitor in production, archive old versions after stable period. Model cards: document model purpose, intended users, out-of-scope uses, evaluation results by demographic, limitations, ethical considerations. Required for EU AI Act compliance for high-risk systems. Automation: CI/CD pipeline triggers evaluation and deployment on new model registration. GitHub Actions: run tests -> compare to baseline -> if improvement, transition to staging -> notify team for approval. MLflow Aliases: add "champion" alias to current best model, "challenger" to candidates being evaluated. Code references aliases, not version numbers, enabling transparent promotion.