The initial concern was that a 1/200 sigmoid scaling might oversaturate the weights. So, the new network was trained with the same parameters and data, but with a doubled K-value for the sigmoid scaling.

To make it work, evaluation-based search parameters have also been slightly increased (≈10%).

Elo   | 4.86 +- 4.23 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=32MB
LLR   | 3.02 (-2.25, 2.89) [0.00, 5.00]
Games | N: 13806 W: 3791 L: 3598 D: 6417
Penta | [251, 1577, 3067, 1744, 264]

For the sake of experiment, increasing the search parameters without changing the network and scale is most likely not an improvement:

Elo   | -0.29 +- 3.55 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=32MB
LLR   | -2.25 (-2.25, 2.89) [0.00, 5.00]
Games | N: 18212 W: 4509 L: 4524 D: 9179
Penta | [229, 2239, 4200, 2194, 244]