Companies generally will train their safety drivers to intervene at any situation that looks dangerous. They tell them not to wait to assure the danger is certain. If the car is not braking for a pedestrian, they don’t keep it going to see if it will hit the pedestrian. The safety driver is told to take over. Later, the team will play the recording back and try to determine what would have happened if the driver had not intervened. Ideally, they will turn the whole situation into a simulator scenario, so that they can test that question in a dynamic way. This is a good approach which many teams do, but less mature teams may not have the resources to do this on every event.
If you do it this way, you can come up with a more interesting number, what can be called “necessary interventions” or at a higher bar, “accident preventing interventions.” You will be very concerned with the latter. The former might include things like a vehicle wandering out of its lane on an empty road — that would not cause an accident, and people are guilty of this all the time — but you would like to know why it happened and fix it if need be.
A large number of interventions should get classed as “cautionary.” The safety driver grabbed the wheel, but the system would have done the right thing if not. These are actually very positive events, since how vehicles perform in dangerous situations is one of the most important things to track.