I think is this a good document and it also brings to the table the threshold images count (around 30) and the alternate neural hash they will keep private (to

I think is this a good document and it also brings to the table the threshold images count (around 30) and the alternate neural hash they will keep private (to guard against adverserial images trying to create a false positive on-device). FWIW, I actually did an amateur threat model analysis in a comment in separate HN thread. I always thought this was called for because the initial document set was just the mathematics, not the people/process/implementation/policy risks and threat model that was the source of widespread concerns.