FinTrust (AWS account 111122223333) runs an automated ML pipeline in SageMaker Studio that must preprocess transaction data, train a fraud detection model, evaluate it, and register a vetted model package in the Model Registry. You need to implement a SageMaker Pipeline that includes a ProcessingStep to run data preprocessing and evaluation, a TrainingStep that consumes the processed data, and a final RegisterModel step that records model metrics produced by the evaluation. Which pipeline implementation approach meets the requirement and ensures the RegisterModel step records the evaluation metrics produced by the ProcessingStep?

The correct approach uses a ProcessingStep to perform preprocessing and evaluation, outputting an evaluation.json file. The TrainingStep then consumes the processed data from the ProcessingStep as its input. Crucially, the RegisterModel step references the trained model artifacts from the TrainingStep (TrainingStep.modeldata) and uses a PropertyFile to point to the evaluation.json generated by the ProcessingStep. This allows the RegisterModel step to correctly associate the evaluation metrics with the registered model package. The other options are incorrect because: Defining a single TrainingStep and calling boto3.sagemaker.createmodel() bypasses the SageMaker Model Registry's structured approach for model versioning and metadata. Using a CreateModelStep immediately after preprocessing and registering the model before training is illogical, as the model isn't trained yet. Omitting RegisterModel and relying on a downstream Lambda adds unnecessary complexity and doesn't integrate the model registration directly into the pipeline.