(This was Richard Jones' idea.)
Erez Petrank gave an interesting seminar about how to make reference-counting garbage collection strategies work efficiently -- essentially by inferring what changes were made by how references change, rather than explicitly updating the counts all the time. However, one of the reasons that updating refcounts is expensive is because you need to avoid parallel usage problems -- and in occam we can prove that these don't exist after the usage checks, so we might be able to get some optimisations out of a conventional refcounting strategy without too much effort.