This is guard rails put on an early release to make it more conservative in some situations and that was likely a situation where it was not sure if it should be conservative or not.I wonder why the five star humans can't drive at a constant speed without a pace car. Maybe it's five stars out of ten.
1. Release it wideHow does "autolabel" know if laws were broken, specifically?
2. Find errors
3. Fix errors one by one
For example if they find that it doesn't stop at stop signs in CA and NHTSA says to them fix this, they ask autolabel for situations in CA where there was a stop sign and the driver didn't stop and remove them and then do the hard work of gathering data of drivers actually obeying the stupid law and adding them. Problem solved, customer are complaining but at least NHTSA are happy.
For the turn on red they first make a map of where the law applies, then ask for situations where driver does this and add them to the dataset. Then it looks for situation where the system fails and asks for even more data of those specific situations where the system behaves correctly.
If they cannot find enough data they ask their hired drivers to go there and manually gather the data, like they did with Chucks unprotected left turn. Fun story, I was once part of team gathering data. We realized we didn't have enough data of the lane for emergency vehicles only. Our professional test driver did 10 laps of driving in it, knowing he was breaking the traffic laws.