So our current master is an interpolation/mix between the network evaluation and our classical evaluation function.
We know that if we remove the NNUE and let stockfish work only on classical we lose around 120 ELO points.
I would like to know how much ELO would we lose if we remove the classical and let stockfish work only on the NNUE.
Do we already have such information?
if not, can we test that on fishtest?
if yes, can we re-run the test to get a more accurate measure?