Risky business requires systems which have measures of risks and hazards.

This is what bothers me about some of the self-driving efforts. You need a metric for "how long until we hit something on current course", and that has to be connected to braking. This has to work even if you don't know what the "something" is or the planning and evasive action subsystem hasn't decided what to do yet. Tesla's "plows into unrecognized stationary objects projecting into left side of road" feature indicates that they relied too much on recognizing known objects. This is also a criticism of automatic driving using deep neural networks. That's likely to perform well until it performs very, very badly.

How would you say, prevent me from accidentally removing more servers than I should using the change boundaries Skyliner provides.

Asking for a friend...