The Easiest Step in the Chain

Team A is reading it correctly. The data, the architecture, the loss function and the noise level are all held fixed, so the target is the only thing that differs, and the lower average squared error belongs to the better model. Team B's target is making an easy job needlessly hard.

The two numbers are scored against different targets, so they cannot be ranked against each other. At step 5 the input is already nearly the answer Team A asked for — a network that copies its input straight through scores about $0.010$ — whereas Team B's target cannot be reached by copying at all.

Noise is random, so it cannot be predicted and Team B's number is mostly measuring luck. Noise prediction only becomes workable deep in the chain, so the sensible design is to train against the image early on and switch to the noise target later.

Knowing the noise gives you the image and knowing the image gives you the noise, so these are one objective written two ways. Two numbers this far apart can only mean that one of the two training loops has a bug in it.