[comnetxr]

They explain in the paper that the quantity plotted in Fig. 4 can also be interpreted as a probability of zero errors of running the circuit. You can see that it goes down quickly with number of qubits and number of cycles, showing the need for either error correction or much improvement in future devices in order to get any useful calculations done. The win here is that the probability of success is non-zero, as they say in the paper: "A single bit or phase flip over the course of the algorithm will completely shuffle the speckle pattern and result in close to 0 fidelity". So the only way to get the right result is to have runs with zero errors.