Matt Fletcher wrote:
then there is an iterative process based on results between ungraded players until the system settles down to consistent grades for everyone?
That's correct. Every so often it seems to converge to implausible values. This is connected with players scoring about 90% or 10%. The 40 point rule seems to have an effect. As an illustration suppose you won nine games out of ten against players rated an average of 130, none of whom are below 110. For an existing player of 150, that's a performance of (1300 + 8*50) /10 ie 170. But if none of the players exceeded 150, the performance of player with a grade of 190 would be (1500 + 8*50) ie 190.
The iterative process does sometimes appear to get confused as to what outcome is expected, particularly when it encounters more than one ungraded player in a results history. Still it might have been fixed behind the scenes

.
In the unlikely event that you wished to explore this further, threads with posts by E Michael White are a good source of material.