For many distributed application, it is still hard to make correct decision while using not up to date knowledge of the system state (network latency, failure...). In load balancing system, there is a difficulty to learn at all instant the system load to assign a processor to a submitted task. For such applications, if the overall system state is known with a certain limited error the distributed decision will be correct. We target applications formed of processes located world-wide. These processes cooperate by sharing a replicated object that represents the system state. The shared object is replicated on each process. This paper describes Sabbarus a distributed plate-form offering replicas divergence control protocol for these applications. We consider processes with well-known profiles; however the actual behaviour of each process could diverge from the profile.