Every recommendation has context.
A card page recommends a talent evaluated with elemental support. A deck page optimizes the talents of all four cards together. The two recommendations can differ without contradicting each other: they answer different questions.
Compare the team benefit.
A generation or support talent may give more to its partners than to its holder. In the workshop, use optimization to test every combination across the four cards, then manually try alternatives that interest you. The model uses one active talent per card.
Check the sample size.
The library ranks talents by average gain across the cards that have them. Some talents exist on only one card. Check their sample size and effect text before generalizing the ranking to another composition.
