What jumps out to me is that the optimizer turns fault injection from a one-shot lab trick into something you can actually budget, which changes how people run evals. If a campaign can go from 5,000 blind attempts to a few dozen tuned ones, then the old "we tried a bunch of glitches and nothing happened" style of negative result gets a lot less meaningful unless you also say how much search was spent. It also makes reproduction the more annoying part, not discovery. Once you have a parameter set that flips AES or control flow, the paper says SOBAS can squeeze more repeatability out of it, which is the bit that matters when you're trying to turn a fluke into a reliable exploit path or a reliable countermeasure test. That means the next downstream benchmark I want to see is not just success rate, but how stable the same setting stays across boards, temperature, and power supply drift, because that's where a lot of these demos fall apart.

it optimizes search, not the physics, so the real bottleneck stays setup.