Variance (left) and bias (right) of the latent mean of each latent parameter (one line per parameter) when white noise is added to a high-S/N spectrum, in units of the latent variance. The latent variance underestimates the uncertainty in latent space for spectra with S/N < 25, which are 90% of the spectra. The latent mean is not significantly biased if the spectrum has S/N ⪆ 8.