> Yet in ways, this paper seems like a confession that he's reached something of a brick wall with the problem.

If "the problem" is that of providing accurate, fully static types for this type of transform, then of course yes. ;)

I think the end result is improved (easier to learn and more flexible) for consciously relaxing that goal, however.

(Edit to muse: I think 'relaxing' is a far more intelligent & interesting approach to overly-ambitious goals, as opposed to 'abandoning'.)