How deck estimates work
Eight supported cards in. A model estimate out. Use it to compare ideas, not predict a guaranteed result.
Deck A
Matchup model
Supported inputs
Estimate
With support notes
Input
Complete decks
Eight distinct cards, forms, levels, and tower.
Search
Bounded candidates
A limited search, not every legal deck.
Validation
57.77% accuracy
On 48,638 held-out matches for this baseline.
Status
Model coverage
Rare decks and balance drift need more review.
What goes into a matchup score
The frozen, calibrated two-layer attention model was trained on recorded Ranked-style matches. Inputs include card identity, Evolution or Hero form, card level, tower troop, and tower level. Deck order does not matter.
The service validates eight distinct base cards and supported special-slot combinations before scoring. The model returns an estimate for a decisive match under supported inputs; it does not observe your starting hand, placements, timing, opponent skill, or decisions.
How counter search chooses candidates
Search starts from observed decks and explores legal card replacements within a fixed budget. It scores complete decks against the target or a frequency-weighted basket of historical decks and returns the strongest supported candidates it found.
Because search is bounded, the result is not a globally optimal deck or a guarantee. Unusual combinations, rare cards, balance updates, card levels, tower troops, and player skill can all affect a real matchup. Support warnings are retained in service responses.
How to interpret the validation result
The two-layer attention model achieved 57.77% accuracy on 48,638 held-out matches in its evaluation. That is a test-set metric for this model version, not a forecast of your personal win rate or proof that every rare matchup is reliable.
Further releases need checks for calibration, rare decks, and drift after balance changes. Use the tools to compare ideas, then make the decision based on your own cards and play.