Precision vs Faithfullness

[apaz-cli]

I was reading the implementation of MXFP4/INT8 dot product for gpt-oss, and it occurred to me. Since it's memory bound, we could get free precision by accumulating in fp64, or using kahan summation, or both, and it wouldn't be any slower. This is probably the case every place there's an accumulate.

Which does llama.cpp care about more, implementing the models faithfully, or getting extra free precision for better outputs?