Comment by slalmeidabbm

This would be the way to go with the initial offering. Adding static code analysis + LLMs will help with reducing LLM usage and hallucinations and then adding a way to test out the policies to make sure that they are enough to run the code without being too broad will increase trust in the results.