Benchmarks
500 beginner boards, 9 × 9, 10 mines. Every agent sees the same seeds from 20000 through 20499.
best win rate91.0%
Strategy CNN hybrid on 500 held-out 10-mine boards. The expert control hybrid reached 35.4% on an independent 500-board confirmation set.
The harderboard check.
The same control and strategy models also ran on 16 × 16 boards with 40 mines. The strategy model adds three playbook-derived channels.
66.0
71.2
74.8
75.4
cnn+5.2 pp
cnn hybrid+0.6 pp
0.26
0.20
1.48
0.47
brier−22.4%
calibration−68.4%
Lower is better for Brier score and expected calibration error.500 held-out intermediate boards.
agentwin rateavg movesdecision rule
random0.0%4.8uniform valid move
local heuristic33.0%10.9clue count
symbolic87.6%19.2constraint proof
exact86.2%18.6frontier enumeration
hybrid90.6%19.7symbolic + exact
strategy CNN88.8%19.2strategy channels
strategy CNN hybrid91.0%19.7model rank + proof