I think the key would be having a technique to discern when approximations were valid & specialize the right code for the right part of the problem. For instanc

I think the key would be having a technique to discern when approximations were valid & specialize the right code for the right part of the problem. For instance, if the program mistakenly tried to run the airframe using DFT instead of FEM, it would never get anywhere, but DFT might be exactly what is needed to model the outcome of cosmic ray bit flips in the nav computer transistors. If the simulation is able to make judgement calls like that, it's pretty much a artificial general intelligence.