What problem does it solve? Machine learning teams lose reproducibility when runs lack dataset versions, models get copied by hand, and promotions happen without written criteria. This Skill turns MLflow from an installed tool into an enforced organizational standard covering logging contracts, naming, promotion gates, and rollback. ## Core Features & Use Cases - Mandatory logging contract: Defines the fields every run must log (hyperparameters, metrics, dataset name and version, git SHA, environment, seed, model artifact) and rejects incomplete runs at review time. - Registry stage workflow: Specifies the None → Staging → Production → Archived lifecycle with written promotion criteria per transition, CI/CD listeners on stage changes, and archive-never-delete governance. - Rollback runbook: Provides a registry-based rollback procedure that pins deployments by model version instead of latest tags. - Use Case: An ML platform team audits recent models in use, finds runs missing dataset versions, then rolls out a shared MLflow tracking server on Kubernetes with Postgres and S3 artifact storage, a run-linter in CI, and promotion checklists per model family. ## Quick Start Use the experiment-registry-standard skill to audit our current MLflow runs against the logging contract and draft the promotion criteria and rollback runbook for our model registry.
