The how-to is definitely the trickiest part, I really like @jengajojo suggestion of doing so through QF, but I also think we have to consider based on which metrics/parameters we would compensate delegates - and given that some metrics might be subjective, it could be difficult to implement this in a fully trustless way. Apart from this, I think that whatever scores are assigned to candidates should be reset after some time, just to prevent candidates that once were very active to keep on receiving incentives after they stop contributing, if that makes sense. This is why we also proposed having governance epochs, which could complement this but is not strictly needed.
So to summarize this, imo we should try to make this as trustless as possible, but also ensure that we can have sybil resistant metrics to base our rewards on.