In the default IsrTask, the variance is currently set after dark subtraction. This means that photon noise from the dark is not included in the variance plane, which is incorrect. The variance should be set after bias subtraction and before dark subtraction.
Hsin-Fang Chiang also points out (
DM-4191) that the AssembleCcdTask with default parameters requires amplifier images with variance planes, even though the variance cannot be set properly until after full-frame bias subtraction. I believe that AssembleCcdTask only requires a variance plane in the amp images because it does an "effective gain" calculation, but I suggest that this isn't very useful (an approximation of an approximation, and you're never going to use that information anyway because it's embedded in the variance plane with better fidelity). I therefore suggest that this effective gain calculation be stripped out and that AssembleCcdTask not require variance planes.