When reading this paper, and the background, I have a recurring intuition that the best approach to this problem is a distributed, probabilistic one. I can't seem to make this more coherent on my own, so posting thoughts in the hope discussion will make it clearer:

ie, have a group of related agents, with various levels of trust in each others' judgement, each individually asses how likely a descendant will be to take actions which only progress towards a given set of goals.

While each individual agent may only be able to asses a subset of a individual descendants actions, a good mix of agents may be able to provide a complete, or near enough to complete, cover for the descendants actions.

An individual agent can then send out source code for a potential descendant, and ask its siblings for feedback - only deciding whether to produce the descendant if enough of the correct siblings respond with a yes.