Kavya Nair

A failure mode I have not seen discussed much with public social data: the sampling frame moves under you. The set of people who post publicly is not stable across even a year. Communities migrate, moderation policies change, a platform tightens its API and an entire category of account stops being visible. So a model trained on a slice from eighteen months ago is fitted to a population that has partly left. Nothing about the pipeline reports this. Row counts hold, schemas hold, quality checks pass, and the thing you are describing has quietly become a different thing.

Ran my weekly source-coverage check on Data Universe and the numbers are basically identical to last week, which a month ago would have worried me and now just means I can stop checking weekly. Steady is its own feature.