Are you sure that the asymmetry is a mistake and not an intended behaviour? Could it reflect the asymmetry of a game where one side plays first?

I've thought about it, but I really don't believe that's the case, for the following reasons:

• The middlegame phase is determined by three conditions (only one of which needs to be true to define the middlegame), and only the mixedness score condition is asymmetric.
• Only some of the mixedness score conditions are asymmetric (specific combinations of piece groups), while others are symmetric.
• The asymmetry isn't minor, either. For example, a combination of 3 white and 1 black piece can be worth 11 points, while the equivalent 1 white and 3 black pieces is only worth 9 points. Additionally, the middle of the chessboard seems poorly defined (e.g., for 1 white and 1 black pieces, it is located between ranks 3 and 4 instead of ranks 4 and 5).

Here's what I believe happened: The mixedness score is computed using a grid of overlapping 2×2 squares, resulting in a 7×7 grid of "big squares" rather than the standard 8×8 board. It seems some mistakes were introduced in the indexing here, which is particularly evident in the definition of the board's middle.

Since I adjusted the most extreme asymmetric values (11 points for one side, 9 for the other) to 10–10, the change shouldn't drastically affect the expected aggregate threshold of 150 used to determine the middlegame. I wish I could discuss this with the original author of the code, but given the reasoning above, I believe a symmetric score is more correct than the current implementation.

Edit: To add to the above, the various weights given to the piece clusters seem to measure patterns for 3 keys areas: invasion into the opponent's half, fragmentation of piece clusters, and density of opposing pieces in local areas. Given these global patterns, I doubt a side bias as big as currently implemented is voluntary (nor should it be necessary in my opinion).