Here $\rho$ controls the guidance strength. The reverse-kernel structure keeps the same base mean and nominal noise scale $\sigma_t$, while the whitening operator uses confidence-interval projections ...