A Pokémon card agent that plays itself

The job is to play Pokémon cards against other people's code. Not the real game on a table. A simulator called cabt, which is a rules engine that only offers legal moves, so the agent never has to guess what is allowed.

An agent here is a small Python file plus a 60-card deck. Each turn the engine asks, which of these legal options do you want, and you return an index. That is the whole interface. I shipped it as a tar.gz named submission.tar.gz.

I did not start by inventing a genius policy. I watched what was already winning. Kaggle posts daily episode logs, thousands of finished games as JSON. A kernel I called ptcg-mine-meta parsed those files, grouped them by archetype, which is the family of deck, like a fire attacker or a stall list, and kept the most common winning deck in each family.

Then I made that deck play itself. The self-play kernel copies the agent into a writable folder, because the Kaggle input disk is read-only, and runs fifty games with sides swapped so going first is not a cheat. I wrote the win rate and the error rate to a JSON file. If the agent crashes, you see it here before you waste a submission.

The ladder is a rating, like chess. My accepted scores went 621, then a dip to 532, then 549, 571, 642, and 708. The last one is the one that counts. 2057 of 6807 teams. Mid pack. The climb is the part I care about, because it means the miner plus the self-play loop actually moved the number.

What I would not claim. I did not solve the game. 708 is not the top of that board. The miner clones decks that already work. The self-play check is a sanity test, not a full search. If you want the next step, it is a better policy than copy the winner, not another scrape of yesterday's log.