There's for sure some minor tuning of the positional weights for the first move to be done, but I've looked at the scores of tutor mode for an example that matc

There's for sure some minor tuning of the positional weights for the first move to be done, but I've looked at the scores of tutor mode for an example that matches the one shown (for a corner and middle equivalent) and the model agrees with me that top middle is superior by 1 p.p. On average for boards where both positions don't appear in the same board over a few samples, the corner was rated an average of 1 p.p. higher but it's not apples to apples when its across different boards as the scores bake in the weight of first movers advantage. I can't be arsed to test this particular board in the example with manual code. There's no such thing as an unbeatable strategy. My point is not that the greedy solution is superior at playing the game. My point is that it will perform with very nearly perfect play with more than 1000 times few calculations. With tuning it could very likely achieve provably perfect play. You could similarly use the logic of this greedy strategy to simply remove obviously incorrect spaces of the dynamic equation to get globally optimal answer with probably > 95% fewer calculations on average. This kind of greedy vs. slow but globally optimal solutions are super common in real world problems.