Two interns improved an ice-rink compressor model on their laptops. Nobody can recreate the run, compare it to last week's version, or roll it toward live alarms. Which practice is missing?

When two interns improve an ice-rink compressor model on laptops and nobody can recreate the run, compare it to last week's version, or roll it toward live alarms, the missing practice is MLOps: repeatable training, tracked experiments, and a governed path to production. MLOps is the operational discipline that turns ad-hoc notebook wins into auditable, reproducible releases—versioned data and code, experiment tracking, evaluation gates, and promotion into the alarm path the facility actually trusts.

Reducing MLOps to a specific Python library name mistakes a concept for a product. Tools help, but MLOps is the set of practices and pipelines, not a single package brand you must memorize as the definition.

Speech synthesis of the training logs is an unrelated generative audio task. Reading logs aloud does not recreate experiments or create a release path to compressor alarms.

Deleting the held-out test slice so the laptop score stays high destroys evaluation integrity. A high unreproducible score without a test set is the opposite of production readiness for live ice-rink alarms.

Exam caveat: experiment tracking without a deployment path is incomplete MLOps, and a deployment path without reproducibility is equally incomplete. The scenario fails on both recreation and promotion.

Operational check: require every candidate compressor model to log dataset version, code commit, hyperparameters, and metrics; block promotion to live alarms until a second machine can reproduce the training job and beat the registered champion on the agreed test protocol.