Technically, you can get the same answer twice from an LLM, but only when you control the full input. When a model is being run, a random seed/hash is applied to the input. If you run the model locally you could force the seed to always be the same so that you would always get the same answer for a given question.

I’m failing to see why the creative writing machine is better than a simulation set to ‘rough’.

This doesn’t sound like they are using LLMs for processing their 3D models. The way it’s described in the article sounds a lot more like some machine learning model trained on physics simulations for aerodynamics.