I think you want to provide feedback for each discrete or continuous action by the user. This is quite easy if we are relying on structured editing, but much harder for free form text. My current policy is to re-execute memoization units after each keystroke, though certain syntax and type errors can prevent the executed expression tree from changing (but then the feedback is a syntax/type error!). Getting live feedback also requires toleration to transient syntax/semantic/and execution errors, and to be fairly efficient. In this case, we should tolerate and react to '1' and '10', even if they lead to a program in a weird state (e.g., x - N, where N should be > 100), but you still want to show some kind of change. And this is where live programming is very different: the steady frame that we are viewing while editing should always CHANGE when we edit, meaning it should be focused on some execution result of the code being editing. Doing something without immediate feedback is being a sad panda.
