Where is AP % confidence explained anywhere? Who actually knew that a bump in a turn could kick it out of AP? I had to discover that myself which is what we're all doing - randomly. Some are not so good at it.
I'm going to draw from experience on a problem I saw at Intel. I gathered human related excursion data x-factory for 3 years to create a pareto of "training" issues on a single operation. Long story short, people knew they weren't supposed to do things, but they were never told why. When reasons were explained, there was a measurable upside in yield as reported by Production. Most Technicians cut corners to save time, but had no clue how this impacted yield as pre/post testing showed. It was not a controlled experiment, and true, they knew people were watching them... but I'm a firm believer in an understanding of things and not just blind rules. The only question should be to balance safety and knowledge against Intellectual Property. But I doubt there's anything new here until Tesla creates its own silicon. It's all OpenAI.