A reward curve which started as n^2 / exponential (thus flat), and then later changed into linear would work against self-voting as well as excessive rewards:

• 'Conciliation' of the reward curves?.

• Or, even better, an idea of @clayop.