I have lots of personal and professional experience with statistics, safety analysis, and risk evaluation, thanks. I literally do it for a living. One of the unique aspects is that I actually have to show my work to regulators, not just wave my hands over it.
It's easy to see that Tesla does not primarily consider the "safety score" as an estimate of your collision risk. It's a unitless, "beta" metric. Look at the
Tesla insurance page: They can't even tell you how the safety score impacts your rates without adding in a completely undefined "safety score from trips," and it's clear how much they focus on mileage as an input and where you live (you'd think Predicted Collision Frequency would already have PREDICTED that...)
Are you really going to defend Tesla not including speeding or red light/stop sign running, yet including false forward collision alerts as a solid statistical model that will really predict their losses and that is PRIMARILY designed to set insurance rates, not filter for FSD "limited access, not wide release" BETA access?
Only time will really tell who is right. Tesla has not used this data for long enough to know if they will be able to keep it. A few months/years of losses from using this number and they'll have to change. Even Tesla screams "BETA" every time they mention it and Elon says it needs to evolve. Like they are already setting up to change it and know it is flawed. You're awfully sure that Tesla is well informed here for a company that slaps BETA on even their insurance offerings and has been doing it for 6 months.