GitOps Wasn’t Built for Models, and It Shows

GitOps works well for stateless services but breaks down when applied to ML models. The core issues are: Git cannot capture the full source of truth for a model (training data, seeds, hardware matter too), green CI/CD pipelines only confirm a container built — not that the model is accurate, and rollback doesn't restore original behavior if the data context has changed. Teams that succeed treat datasets and training runs as first-class versioned artifacts, include real model evaluation as a quality gate, and define rollback as restoring both the model and its data context — not just the container image.