Build a Supply-Chain Checklist for Adopting a Third-Party Model

A product team wants to download a popular open-weights image classifier from a public model hub, fine-tune it on customer data and deploy it in a service with access to a production database. The model card looks professional and has thousands of downloads. The team lead says: "It's from a well-known hub, a scanner already checked it, so we can just load it in the training notebook."

1. Explain concretely what is wrong with "just load it", including what a pickled model file can do at load time.
2. Write the checklist you would require, grouped by stage.
3. Which risks remain after a clean file format and a clean scan?
4. How would you test the model for a backdoor, and what are the limits of that test?