0. The answer in five seconds
First, let AI play a huge number of games and compare many deck builds. Then take the narrowed-down 40 cards and a clear plan into Ranked. There you cut the cards and moves you don’t need. When you meet the same opponent again and again, you start to remember that person’s habits too.
AI: cuts down the search space
You: cut down the choices in real games
Rematches: cut down the guesswork about one opponent
Your brain: stores whole board positions
In this anonymous run, Aggro Nightmare reached Diamond Group and kept winning a lot afterward. But the most interesting part was not “Diamond is strong.” The more rematches there were, the more the game turned into a long series that only one side knew about.
Against Witch, I eventually did not even need to remember the full card names.
Researcher guy. Experiments. Lion. Okay, next.
The calculation did not disappear. Old calculations were squeezed into a small package, so the answer comes to mind the moment I see the board.
1. AI walked through the minefield first: the human did not search for 40 cards from zero
Before this deck, there was a long stretch of AI research. It started from official tournament decks and covered a card database, a rules engine, mulligans, going first or second, play policies, the lowest win rate against each matchup, and downside risk. On the full-rules side it ran up to 12,480 games, and it hit a trap along the way: even when the rules are coded correctly, weak AI decisions can make a hard deck look worse than it is.
On the first day of Set 9, it compared 32 builds over 2,324 games. Inside that model, Aggro Nightmare came first with an average win rate of 81.0% and a worst matchup of 63.9%. For Pirate Swordcraft only, it tested 5,904 games plus 1,944 games on unseen seeds, and the baseline 40 cards scored 91.4%.
Neither number means you will win 81% or 91.4% of your games in Ranked. What matters more than the numbers is that a huge number of candidates were thrown out, so the human had a much smaller range to try.
Before a human plays 100 games and thinks “is this weak?”, the machine steps on the mines first. The human starts by checking the answers.
2. Diamond and rematches: it is a single game on screen, but a best-of-seven in your head
In Shadowverse: Worlds Beyond, Group is separate from Rank. Officially, it goes up and down based on your recent wins and losses. Normal BP and MP do not drop when you lose, but Group can go down. Diamond is the top Group, and when you move up to Grand Master, the official rules start you at CR 1600.
So reaching Diamond shows more than “I played a lot.” It is a sign that you had been winning quite a bit recently.
On the other hand, we cannot say that Diamond always means fewer opponents, or that you only meet other Diamond players. The matchmaking formula is not public.
Even so, in real play I met the same opponents again at quiet times, such as weekday afternoons and late at night.
Game 1: an ordinary single game.
Game 2: you know something from last time.
Game 3: you start guessing that person’s habits.
On screen it is a single game. In your head it is a best-of-seven.
“This person always answers this way on turn 5.” “They don’t hold back here.” “On a board like this, they go for removal.” Before you notice, you are no longer countering a deck. You are countering that person.
3. This Experiment Witch can be remembered as “researcher + experiments + lion”
The Set 9 Experiment Witch builds up experiments with cards like Sephie, Sinner of Mystic Arts, Entranced Researcher and Immersed Experiment. Immersed Experiment can affect the board from early on. Once its condition is met, it becomes about 5/5, and once Sephie’s setup is in place, it brings Storm pressure too.
From my side, the warning signs are fewer than the full card text.
A researcher shows up
→ experiments pile up
→ 5/5 range
→ Storm range
The “lion” here is a nickname based on how the experiment artwork looks. It is not the card Lio, Manifestation of Exhaustion.
Ranked is not a spelling test of card names. If “the lion showed up = play this answer” is the right branch, that is enough for real games.
On the Aggro Nightmare side, the plan got stronger too. Around turns 5 and 6, don’t carelessly drop extra followers that would give the opponent healing or board recovery. Instead, gather the cards that reach the face, and save your Storm for one big push.
The goal is not to take the board back. If you remove the enemy leader, the whole board disappears with it.
4. What “I’m not calculating” really is: chunking, or saving the answers
At first, you think about PP, the number of experiments, healing targets, the next Storm turn and exact lethal one by one.
When you meet the same archetype many times, and even the same person, all of that joins into one lump. It is close to what cognitive science calls chunking.
To a beginner, it looks like this:
Card A + board B + remaining PP + guess at the opponent’s hand + next turn
To someone used to it, it ends with:
“Ah, the lion shape.”
You did not throw away your thinking. You saved the result of earlier calculations in a form you don’t have to redo every time. You didn’t break the calculator. You saved the answers.
On top of that, Aggro Nightmare’s own decision tree is small right now.
If you can chip, go face.
If it would feed their healing, don’t play it.
Save Storm and use it together.
Only stop a guaranteed lethal.
If you use VS Zeta, check your own life too.
Because your own side is simple, you can spend the spare brain power on your opponent’s habits.
5. I win a lot. But keep it separate from the “true win rate”
The high win rate I feel now mixes at least five things.
- The raw deck power of Aggro Nightmare
- The effect of narrowing the first candidates with AI
- The effect of cutting unneeded cards and moves in real games
- Local adaptation to common matchups such as Experiment Witch
- The effect of learning one person’s habits through rematches
Winning 70% against 100 different people is a different thing, statistically, from winning 70% after meeting the same small group many times and learning them. The second is also a skill that wins in Ranked, but it is not the deck’s general strength.
To check how general it is, take a new sample of at least 20 to 30 games at a busy time of day, against different opponents. If you still win at a similar rate, it was not just “memorizing the late-night villagers.”
But if all you want is to win in Ranked right now, memorizing the late-night villagers is a perfectly legal way to improve.
6. Final conclusion: hand the searching to AI, and the human remembers only the exceptions
The whole thing is one line.
AI research
→ cut down the candidates
→ play Aggro Nightmare in real games
→ lean toward quick damage instead of board recovery
→ learn the standard answers to Witch
→ reach Diamond
→ learn the same opponents’ habits
→ board positions become chunks and long thinking disappears
At the start, it was:
“Card game math is a pain, so let AI do it.”
At the end, it became:
AI: “I watched ten thousand games.”
Human: “The lion showed up. Okay, this pattern.”
This is not a story about AI playing the game in place of a human.
It is a story about handing off the searching, so the human can focus only on the exceptions in the live game.
