See point #5 under notes;
Since it cant simulate the action of the user and it can only simulate its own actions, it would essentially do nothing. The only action the box can do is give its output, so when the user fires up the machine the only thing the box can simulate is itself giving the output. So the simulation in the machine is unchanging (nothing happens), because there is no input for the simulator to run. (the Machine can only simulate itself and the only thing the machine does is give output [to dig a little deeper: the machine cant even simulate its own output on the simulation, as it cant simulate something to input the output it generates])
TL:DR the machine is an observer only and can only give advice, it can not effect change on the system. (It is one half of an input/output statement)
The machine will sit there doing nothing until it hits it programmed timeout length of a few decades, then it goes through its simulated output results and compares this to the simulation. It sees that its output had no effect on the simulation (the machine doesn't know that its outputs where never put into the simulation, due to limitations placed on it by its creators) So it thinks everything is about as good as it is going to get, the last output to improve the system would be to "Destroy the machine" and free up its resources.
The reason you get the Destroy Machine output is: a machine is only as good as its makers and will only function as good as its user. OP's machine has a few fundamental flaws, it was never going to work.
** It think the answer OP was looking for is actually hinged on point 5. The machine only factors in the machine output into the simulation. This leads the simulation to an equilibrium as the orderly output of the machine balances out the chaotic human portion of the simulation. Because the machine can only give an (a single) output it gives the last output it tweaked the simulation with "Destroy the Machine".
The reason you get the the output "Destroy the Machine" is that the machine is limited to one output that takes several decades to get to. That last output is always "destroy the machine" as there is no longer a function for the machine. To fix this this, (assume that processing time of the simulation and the real world are at the ratio 2:1 [anything around the 1.5:1 and below would be useless, as majority of the information will be out of date]) just redesign the machine to allow it to output 140 outputs. [ 1/4 - 1/2 of those outputs will be useless as their corresponding time period would have passed and you wouldn't be able to act on them (Time in the real world still ticks along while the simulation is running)]. This will get you an output that has "destroy the machine" in it and not one that is wholey "destroy the machine"
*** Actually the more i look at the information provided, the more holes i see appearing in the ability of this machine to actually work. The 5 note points, describe the machine as non-functioning. Sure the machine is on point with the science, processing and all that stuff, but the simulation never progresses that far, due to the Machine only being an observer and only giving advice output, which needs to be actioned by a human in real life and in the simulation, except the machine can only simulate itself, nothing else. Kinda like forgetting the password to your PC, the PC still functions and all that, you just dont have access to it.