I'm curious, CRDTs are good for having data converge in the same state but not necessarily ensuring that the state it converges to is "correct". How does this solve that?
You can't guarantee "correctness" (if you define that as zero information loss and/or external causal consistency) in any multi-leader system partly because of unsynchronized clocks and partly because, well, you have to deal with the conflict somehow. You can think of CRDT as a tool to minimize relevant information loss (ideally to zero) when two concurrent writes are converged.
