Generalization

Cards change; their underlying effects recur. A model built from mechanical primitives can carry knowledge across formats, even when a card has no trained identity.

Effects recur across formats

The report connects older archetypes to modern ones by effect similarity. These mappings explain the transfer hypothesis; they are not a controlled benchmark of all four matchups.

Shared effects let learned decisions transfer to unseen cards.

A measured test: Mew World Order

This separate experiment uses 60 cards from 1999–2018, none in the competition pool. After adding their rules to the engine, we tested the submitted model without training on them. Both seats played the same list against a random-action opponent. That tests transfer to unfamiliar cards, not competitive strength against expert play.

zero-shot 37–23 fine-tuned 95–25 vs random

cards spanning 20 years of the game, 1999 to 2018

of them in the competition pool

new mechanics the engine grew to play it

62% → 79%

against random play: the submitted model unchanged, then after fine-tuning on 3,600 searched games

Zero-shot, then fine-tuned

vs. random, seat-paired

37–23

61.7% over 60 games, zero-shot.

Fine-tuned, screen

47–13

78.3% against the parent's 61.7%: a paired gain of +16.7%, 95% interval +2.5% to +30.0%.

Fine-tuned, confirmation

95–25

79.2% against 64.2% on fresh seeds no run had touched: +15.0%, 95% interval +4.6% to +25.0%.

The zero-shot result and the fine-tuned results answer different questions. Only the first uses an unchanged model; the later runs train on 3,600 searched Mew World Order games. The confirmation uses fresh seeds, with paired comparisons reported above.

Explore the representation in the card space explorer, where these cards are marked MWO.

The experimental deck: 60 cards, 20 years