Retraining is a new candidate, not an automatic upgrade.
Understand why fresh outcomes can justify another training run, what to compare between model versions, and why deployment remains a separate review step.
New outcomes give you more examples to learn from. They do not guarantee that the next model will be better. Treat retraining as another experiment with an explicit review before deployment.
Check what changed in the data
Before evaluating a new candidate, understand the latest examples. Have labels become more consistent? Has the source changed? Are the new cases representative of the questions the model will face?
A larger dataset can still contain incomplete outcomes or inputs that reveal the answer. Revisit the task definition and review the prepared examples whenever the source or business process changes.
Compare versions with a clear question
Decide which behavior you want to improve. Compare candidates on a consistent evaluation set where possible, and inspect recent cases separately when they represent a new situation. Look at the mistakes as well as the summary results.
Keep track of the data and settings used for each experiment. If the task definition changes, make that visible in the comparison rather than treating the numbers as directly interchangeable.
Keep deployment explicit
Tuned Predictions supports scheduled refreshes and retraining candidates. A new training run does not automatically replace the deployed model. Evaluation and deployment are separate steps, so you can review the new version before choosing to use it.
After deployment, use traces to inspect the version and inputs behind a prediction. As real outcomes arrive, they can inform the next review. Predictions remain estimates, and you decide how to act on them.
Explore the product workflow or contact the team to discuss your setup.