3 ms·
I'm not fully following your logic, but to resolve that problem you need another consensus algorithm over the votes of representatives. And I don't know how we
by dandanua 6y ago
I'm not fully following your logic, but to resolve that problem you need another consensus algorithm over the votes of representatives. And I don't know how we can effectively use the fact, that a probability of harmful representative behavior is very small.
- a1369209993 6y agoBasically, a representative who splits the network by commiting two conflicting votes is (AIUI) by definition defective, and any representative that observes both potential blocks knows that, and can propagate that knowledge to any representative it communicates with (which must be most of them in order for them know what that (observing) representative's vote was). And once most representatives know that a vote was spoiled, they can throw it out (along with the two blocks it distiguishes between, although I think those still need to stored, just the transactions in them are treated as not having happened). But, as I said, I'm not sure about the logistics/provable-correctness of making sure that the half of the network that voted one way and the half that voted the other way always have a chance to compare notes soon enough to catch a spoiled vote.