Notable phantom braking is like a 1 in 10K mile event at least. Meaning genuine hard braking for something that's not needed. Running a red light has to be a 1 in 100K mile even at least. In a geofenced area I'd think it happens less than that.
That puts the odds of those two things happening at like 1 in 40 million
Even when tesla built the cybercab they had 4000 vehicles for testing
That's not the same thing as running the software on millions of cars
An example of this.
If you want examples of a car overturned at night on the freeway, with no lights on, no cars around, no police, and possibly people walking around that you can kill. You need billions of miles of driving in the middle of the night to get sufficient examples of this.
paying a driver to test during these hours is wasteful because you can't test a lot when there's not much interesting data to collect
Only way to collect sufficient examples of this and validate your system works is to test on millions of consumers