AI lifecycle · rule ai-training-without-tracking
Model training with no experiment tracking
A Metacenta review checks this under the rule Model training records what produced which model. Everything below applies whether or not you ever commission one.
What this rule checks
Where a project declares a model-training library, this rule checks that an experiment-tracking library such as MLflow or Weights & Biases is declared alongside it. It reports once for the project.
Why it matters
When a model in production behaves oddly, the first question is what data and which settings produced it. Without a record, the honest answer is usually that the run cannot be reproduced.
How to fix it
Record which run produced which model. Add an experiment tracker around the training loop, recording the data version, parameters and resulting model for every run.
When it is fine to leave
Runs tracked somewhere the manifest cannot show, such as a hosted MLflow server, a separate repository or your own harness, are tracked in practice. We report this at medium confidence for that reason.
What we need to check it
The dependency manifests in your repository. We reason from what they declare, never from what the team does, so the finding is medium confidence at most.