I’m planning to run a DeepSeek model locally to work with private notes. The weights and inference server would be on my machine, but I would still be using a separate chat interface.

That distinction bothers me: local inference doesn’t automatically tell me what the interface, extensions, or update checks might send over the network.

Before adding any sensitive material, I want to test the full setup using dummy prompts. My questions are:

• Is disconnecting the internet a useful first test, or could it miss data queued for later transmission?
• How can I attribute outbound connections to the chat interface versus the model server?
• What should I check for remote fallback, telemetry, and optional search features?

For people who have done local AI network monitoring, what evidence would you consider convincing? I’m trying to verify the surrounding software, not just confirm that the model can generate a response without internet access.