0. Five-second conclusion
AI did the expensive search first. The player entered Ranked with a narrowed 40-card plan, removed low-value cards and lines through real games, then learned individual habits when the same opponents reappeared.
AI shrinks the search space. The player shrinks the decision tree. Rematches shrink uncertainty. The brain stores whole positions as patterns.
The run reached Diamond and kept winning frequently. Against Witch, full name recall eventually became unnecessary: researcher, experiments, lion — same branch again. Calculation did not vanish; it was compressed into recognition.
1. AI walked through the minefield first
Earlier research used tournament lists, a card database, rules, mulligan and play policies, matchup floors and downside risk. A full-rules phase ran 12,480 games and even exposed the difference between correct rules and poor AI decisions.
For Set 9, 32 supported builds were compared over 2,324 engine games; Aggro Nightmare ranked first inside that model at 81.0% average and 63.9% worst matchup. A Pirate Swordcraft-specific study ran 5,904 games with a 1,944-game unseen-seed holdout and produced 91.4% for one baseline inside that narrow corpus.
Neither number is a literal ladder win rate. The value was candidate reduction: AI paid the exploration cost so the human could start at validation.
2. Diamond plus rematches: the UI says BO1, the brain says BO7
Group is separate from Rank and officially moves with recent wins and losses. Diamond is the highest Group; reaching Grand Master from Diamond starts Class Rating at 1600 under official rules.
That makes Diamond evidence of strong recent results, not just play volume. But the exact matchmaking formula is not public, so Diamond cannot be claimed to automatically create a tiny opponent pool.
Still, repeated opponents were observed in some quieter periods. Game one is ordinary BO1. Game two has history. Game three contains hypotheses about the person. The UI still says BO1; the brain is playing BO7.
3. Current Experiment Witch can be remembered as “researcher + experiments + lion”
Set 9 Experiment Witch builds around Sephie, researchers and Immersed Experiment, accumulating experiments that become larger threats and eventually create Storm pressure.
From the opponent side, the warnings compress nicely: researcher → experiment count → 5/5 zone → Storm zone.
The “lion” in this article is not the separate Witch card Lio. It is only a visual nickname for experiment-related artwork in the current Experiment Witch package.
Ranked is not a spelling exam. If “lion appeared” correctly recalls the next response, the label has done its job.
4. “I barely calculate anymore” is chunking, not the absence of thought
At first the player consciously evaluates PP, experiment count, healing targets, Storm timing and future lethal. Repetition merges those details into a familiar chunk.
A beginner sees card A + board B + PP + likely hand. An adapted player sees “the lion position.”
Previous calculations were cached. They no longer need to be recomputed every time.
Aggro Nightmare helps because its own decision tree can be small: face when damage matters, do not donate healing targets, bundle Storm, stop only guaranteed lethal, and track your own life for VS Zeta.
5. Winning a lot is not automatically the universal true win rate
The current performance mixes raw deck power, AI preselection, human tuning, local metagame fit, and opponent-specific learning from rematches.
Seventy percent against one hundred unique opponents is not the same experiment as seventy percent against a small pool whose habits are already learned. To test generalization, take a fresh 20–30+ game sample in a busier period with more unique opponents.
For winning the current ladder, however, memorizing the “late-night villagers” is completely legal optimization.
6. Final conclusion: let AI explore and let the human memorize exceptions
The full loop is AI research → candidate reduction → fast ladder trials → remove low-value decisions → matchup templates → opponent templates → chunking.
It started as: “Card-game arithmetic is annoying; make AI do the expensive part.”
It ended as:
AI: “I reviewed thousands of games.”
Human: “Lion appeared. This branch.”
AI did not replace the player. Exploration was outsourced so human attention could be reserved for live exceptions.
