During writes with replication, multiple copies of the data are sent to multiple volume servers, and the operation will only return successfully if all copies are written successfully.
The reads are actually implemented by the read client. Most commonly, after a volumeId lookup, a list of all the data locations are returned. Usually you just use one of the data locations to read or compose a URL. There are no data consensus checking.
Technically you can implement more checking during reads. But it is not necessary, since for consistency R + W > N (read replica count + write replica count > replication factor), the read count is always 1, the write count is always N, and 1+N is always greater than N.
Chris