A stronger Advanced opponent.
Advanced now compares its own action sequences with possible replies from the visible opposing board. It keeps defensive deployment choices in the search, considers Challenges against Constructs, and can reinforce a blocker that would otherwise be defeated before the opponent Explores.
Same cards. Better decisions.
Beginner and Intermediate retain their existing policies. Advanced uses the same legal actions, resources and card pool. It cannot see the opponent’s hand or either deck’s shuffle order. The reply model uses public characters, Constructs and announced plans, not invented hidden cards.
The search is bounded for browser play. It does not solve every turn or predict every counter.
Fresh-seed checks
144 fresh-seed games: 72 current Advanced versus frozen previous Advanced, and 72 current Advanced versus unchanged Intermediate. Each comparison includes nine faction mirrors with first Companions and nine rotated cross-faction matchups with alternate Companions; policy seat and starting seat independently crossed. Fixed starter decks, same rules and resources. Separate from earlier pilot and tuning seeds.
Small per-matchup samples. Bot diagnostics, not human win rates or faction balance rankings. Advanced uses a bounded public-board reply model and cannot predict hidden opposing cards. The 90% adjacent-difficulty goal remains a target, not a guarantee.
Average decision time in the automated runtime: 32 ms; slowest sampled decision: 535 ms. These are test-machine measurements, not an on-device speed guarantee.
Current results summary · Previous AI’s historical difficulty tests
New rare card designs are excluded from these tests and all current playable decks.