Damn.... I created a branch ran a few tests (because i am not so good in guessing ;) ) and...
Thanks for the suggestion—I tested the released V6 checkpoint with a “win now / block now” override against the same rule with a centre-first fallback.
From 200 fresh 12-move openings, played twice with colours swapped, the hybrid scored 171 wins, 209 losses and 20 draws: 45.25% over 400 games. Neither player used random moves after the opening. The starting boards and their mirrors were excluded if they appeared in the training corpora or the earlier tactical dataset.
On this sample, the model’s choices in quiet positions did not improve on the centre-first fallback. The result doesn’t prove the model learned no positional patterns, but it doesn’t demonstrate a useful advantage over the one-ply baseline.
Separately, on training-board-disjoint positions without an immediate win or threat, the model agreed with the time-limited teacher on 35.7% versus 43.1% for centre-first. That’s not a direct comparison with tactical accuracy, which we measured differently.
The test and results are in commit 8fe1cab; I’ll push it shortly.
