I’ve downloaded a DeepSeek model to run through a desktop chat application. The app connects to an inference server on localhost, and generation works after I disconnect the internet. That’s encouraging, but it doesn’t answer every privacy question I have.

The setup includes the model runner, the chat interface, and optional extensions. I want to check the whole stack before using private documents—not just assume that local inference means every component stays offline.

What would a practical verification look like?

I can use dummy prompts and monitor outbound connections. What I’m less sure about is how to distinguish update checks and model downloads from telemetry that might include conversation content.

• Should I test startup, document import, generation, and shutdown separately?
• How would I check whether anything queues data while offline and sends it after reconnecting?
• Is a process-level outbound firewall rule enough, or could a helper process make the connection?

I’m looking for a repeatable approach to local AI network monitoring, with clear limits on what the results prove. Which logs and application settings would you inspect alongside a packet capture?