Originally, I think EK's plan was to have all nodes (the Core Server) run the validation logic; however, the algorithms written in ElasticPL can potentially take quite a bit of time and memory to solve (i.e. seconds...depending on the complexity), so the validation logic could easily be used to attack the network (i.e. submit a complex job and throw 100's of solutions to the nodes to validate).
Right, this is one of the reasons for the small/subtle changes in the payouts model. Under the model I propose, the attacker would need to re-submit his complex job after each solution he issued to himself. (His job would also have to be "actually" complex, and not just be full of "meaningless" busy-loops which do not contribute to output value, mind you.) This creates an opportunity at each solution found for miners to reject his new jobs (or up tx fee for them) if they feel the network becomes too loaded. At that point the problem becomes no different from an attacker trying to stuff up the btc network with dust transactions filling up the block limit - he can still try to do it but miners can just make it ever more costly for him until he is broke and must just stop because he can no longer afford it.