With evidence
Experiment tracking
Versions, params, artefacts. Without them you cannot compare two runs.
Simple explanation
If you do not know which model, params, and data produced a result, you cannot reproduce it.
Technical explanation
Log params, metrics, artefacts, code versions. MLflow is a tool, not a magic skill. The point is comparability.
Evidence
Public repo anomalie_eth (on-chain anomaly detection + MLflow). Training / exploration, not a Lab product. No hosted MLflow dashboard here.