Our "tank problem" math is way off because a ton of applenets got reused.
FWIW, I basically trust the 90,000 number.
I've been working with a new idea of how to estimate this number more accurately.
If you graph AppleNet# vs Date, you'll see that there are bands of time where the AppleNets strictly increase, and then at some point, the starting number seems to "reset" to a different random value and continue up from there. The duplicate AppleNet numbers we've found appear in different "bands", so it appears they overlapped each other from time to time.
I took the dataset and divided it up into these "monotonically increasing bands" (where the AppleNet numbers strictly increase), performed a tank problem calculation for each band, and added them all up.
This analysis produces a total Lisa production count estimate of 52,876 - which is more in line with what we're seeing than ~10k when calculated over the entire dataset.
Do note that some bands have only a few AppleNets in them, and when you add more numbers to the middle of a band the tank problem estimate tends to decrease, so this likely tends to be an overestimate and is subject to change as we discover new machines.I'll keep adding new machines as I find them and see if the number changes dramatically. Food for thought for anyone else interested in this very specific statistics problem