Two figures decide whether a property gets modelled data, and they pull in opposite directions: one counts daily users whose storage was granted, the other counts daily events whose storage was denied. Raising the consent rate helps the first and hurts the second, which is why modelling has an upper bound as well as a lower one. The same inputs also show how many rows of a breakdown thresholding is likely to remove.
Google does not publish the row minimum that triggers thresholding, so it is an input here rather than a constant. Fifty is a commonly reported working assumption, not a documented figure - the row counts below inherit whatever value is entered.