Design the CI/CD Pipeline for a New Fraud-Detection Repo

You're setting up CI/CD from scratch for a new repo that contains a feature pipeline and a training job for a fraud-detection model, currently trained ad hoc by a data scientist running notebooks locally and manually copying a serialized model file to a shared S3 path that the serving service reads from.

1. List the CI stages you would add, in the order they should run, and what each one gates.
2. Explain specifically what changes about the "manually copy the model file to S3" step, and why leaving it as-is undermines everything else you built.
3. A stakeholder asks "can't we just run the full production-scale training job on every PR, so the gate uses real numbers instead of an approximation?" Explain the trade-off and what you'd propose instead.