Yes to support HA we need isolate Hardware as well as software. It is kind of Failover (in Diameter Protocol).
The high availability feature is strictly an active/standby scheme. One server will be in Active state and other will be on standby. When first will fall then the calls will go to other server.
They can come up with these figures with help of system testing. Just you can say a hit and trial method. eg. Not in real time network but in Testing setup. You will send 500 messages per second by any kind of traffic generation tool which can send the relative messages. you will run it for 12 hours do the HA flip-flop in between and at last just see the results.
How many messages has been reached there and out of that which has gone missed. How many has got reply. This is how you can estimate that the quality of the feature can give this much accuracy and reliability for real time messages.
In real time environment also if the results are same then it is good for the product but in reality the testing scores and real time scores will have minor difference.