In a cloud ML environment, data scientists, platform engineers and outside contractors all have write access to the training data store, the pipeline definitions and the production model registry. Which change best protects the integrity of the models that reach production?

Protecting model integrity requires controlling who can change the inputs and artifacts that determine production behavior. Least-privilege access means data scientists can experiment in development, engineers maintain pipeline code through reviewed changes, and contractors receive only the narrowly scoped access their engagement needs. Separating development from production, with the production registry writable only by an approved, automated pipeline after review and testing, removes direct human write paths into the most sensitive environment and creates enforced checkpoints. An acceptable use policy sets expectations and supports discipline, but it is an administrative control that does not technically prevent an accidental or malicious change. Encrypting the registry protects confidentiality; anyone with write rights can still overwrite a model, and decryption rights do not govern who can publish. Detailed logging with after-the-fact review is a detective control: it helps discover and attribute a bad change but leaves the window open for a tampered model to serve production traffic in the meantime. Exam caveat: preventive access controls come before detective controls when the question asks how to protect integrity. Operational check: list every identity with write permission on the production registry and confirm only the deployment pipeline's workload identity remains.