Demis is toothless, it will not work because you cannot trust people to self censor when their future work at the committee will be at stake.
you cannot setup reasonable bonds in unlimited problem space because there is no example or metric to judge against.
possible solution would be to limit higher priority access to agi to only people that demonstrated integrity in their decisions. all other metrics can be easily forged, while integrity can be publicly verified by agi itself.
it would probably separate agi from politics while allowing it to do governance - agi will do its lowly ai tasks while human meatbag will do politics.