Dark Souls can be completed on RL1 because mechanical mastery supercedes the need for numerical advantage. But in your game, it seems like you need to have a) mechanical mastery, AND b) numerical advantage in order to win.
When you died repeatedly to the Asylum Demon in DS1, did you need to come back with more levels? No; it was a lesson that not all battles need be tackled head-on.
When you lose in NeuralNX, what’s the lesson? If the AI is just going to crush you optimally each time, I suspect it’s not the same challenge as Dark Souls, and it will feel distinctly unfair. If you didn’t even see their kit and abilities, do players even have the tools to learn?
I recently released an abstract strategy game with a powerful AI, too. It started out with pure minimax and topographically sorted heuristic factors. This squeezed the fun out of the game as it was way, way too hard.
I eventually switched to a weights-based heuristic so I could use softmax over the results. This produced N different-feeling variations of the AI, and let me variably weaken by nudging tau.
What kind of knobs does your AI have? In my experience, full blast mechanically optimal AI is distinctly unfun.