I think manual review might actually be the worst way to solve this problem, because you're relying on human labor to fix every single hiccup, which ruins the p

I think manual review might actually be the worst way to solve this problem, because you're relying on human labor to fix every single hiccup, which ruins the point of an automated system anyways. I think for sure they run the route through a bunch of simulations to tune the ML mode, but I seriously doubt they would go as low-level as using telemetry to disable certain features by hand. The best solution would be updating the model so that it makes the right calls on the right type of road - so that when it encounters a similar road, you don't need to wait for another fatal accident before kicking the review process into gear and making the necessary changes again. tl;dr it doesn't scale and would be a logistical nightmare to maintain individual configurations for every road the system struggles on