Honestly, I think the easiest way to do what you’re asking is to solve alignment, build an aligned AGI, and ask for it to help you.
Partially, I think this because I think alignment is actually a lot more tractable than most people assume. The other reason is that “build real Dath Ilan, but real and stable” seems like a strictly harder version of the alignment problem.
To build a stable version of Dath Ilan, I think you (roughly) need to establish a environment of microeconomic incentive gradients that robustly select for agents that are cooperative, powerful, friendly and rational. I think this is a similar problem to establishing literal gradients that select for the same in the circuits of an ML system.
However, doing so in a real economy is much harder because you have less control over a real economy and because other people can interfere with your efforts.