An organization deploys Azure AI Content Safety to moderate user inputs to a consumer-facing AI application. The content safety API returns categories including Hate, Violence, Sexual, and SelfHarm with severity levels 0-6. The security team must configure the application to block any content with Violence severity 4 or higher while allowing Violence severity 0-2 for general users, but allow all violence categories up to severity 6 for verified security researchers. How should this be implemented?

B is correct because Azure AI Content Safety returns severity scores for each category in the API response; the threshold decision (block/allow) can be implemented in the calling application code rather than in the Content Safety service itself. The application can check the authenticated user's role (security researcher vs. general user) and apply the appropriate severity threshold to the Violence category score returned by Content Safety, allowing differentiated policy enforcement from a single Content Safety resource. A is incorrect because creating two separate Content Safety resources for different user groups is wasteful and complex to manage when application-layer threshold logic achieves the same result. C is incorrect because APIM policies can route and transform requests but cannot easily implement per-user-role content safety threshold logic based on AI safety scores returned in responses. D is incorrect because Azure OpenAI content filtering applies at the model deployment level for all users of that deployment; per-user differentiated filtering requires application-layer logic.