The learner searched 48 positions per move before storing what it learned from. I raised it to 200 and let it run a day.
Per cycle it worked. With the deeper search the network’s own first guess survives only 60% of the time instead of 83% — the target it trains on is genuinely better than the guess that produced it, which is the whole mechanism.
Per hour it did nothing. Measured the same way both times, the champion of the moment against a frozen ancestor over 400 games: 12.0 Elo an hour before, 12.1 after. Better targets, four times fewer cycles, and they cancel exactly.
I kept it anyway — 252 Elo earned from a higher starting point is worth more than the same number earned lower. That is a reason, not a measurement, and it belongs in a different column.