The particular case of Box is safe, and is along the lines of special knowledge about Box that is assumed by the type checker, but the special-cased nature of Box has been gradually decreasing over time. I don't think there would be an easy safe way to allow user-defined types to establish same guarantee.

In retrospect, changing the evaluation order of a function with respect to its arguments might have been a better choice, but something like the solution I mentioned is probably the only backwards-compatible option.