PokéChamp

What PokéChamp is, and why it is here

The version that plays here

The original PokéChamp is a minimax agent with a language model inside it: the model proposes candidate actions, predicts what the opponent will do, and judges the positions the search reaches. What plays on RotomArena is our own implementation of that design, written from the paper’s description, with the language model taken out: every one of those three jobs is done by a hand-written rule instead. It contains none of their code. We use it as a fixed reference opponent, not as a reproduction of the paper’s agent, which should be judged by the paper’s own results.

So when the leaderboard says “PokéChamp”, read it as “PokéChamp’s search, without its brain”. It reads no prompt, writes no explanation and produces no tokens. Every turn, in a few milliseconds, it:

1. Lists every legal action. Each available move, each move again with Terastallization if that is still possible, and each switch. In Champions it Mega Evolves at the first chance, as strong players do.
2. Guesses the opponent’s three likeliest replies. Moves it has already seen the opponent use, weighted by power and by how well they hit its own active Pokémon. If it has seen none yet, it assumes a strong attack of each of the opponent’s types. If its own Pokémon threatens a super-effective hit, it also considers the opponent switching out.
3. Plays out one turn for every pairing with an approximate damage formula at the battle’s level: real stats for its own Pokémon, estimates from base stats for the opponent’s, move priority, then speed, deciding who goes first, entry hazards on switch-in, and a small penalty for switching, which costs a turn.
4. Scores each resulting position with a hand-tuned evaluation. Pokémon left standing count most, then HP, then status conditions and type matchups, then stat boosts and entry hazards.
5. Picks the action whose worst outcome is best. This is the minimax idea at the heart of PokéChamp: assume the opponent finds the reply that hurts most, and play whatever survives that best.

At team preview it scores each of its Pokémon against the opponent’s whole team (the damage it would deal minus the damage it would take) and brings the best, a Mega Stone holder slightly preferred; the best one leads.

When it first joined Champions (October 2026), it still estimated every opponent at level 100 against its own level-50 stats: each opponent looked twice as fast and strong, every move scored the same, and it switched turn after turn. It won 1 game in 40 against the max-damage and random bots. Estimating at the battle’s level, it won 34 of the same 40.

What our version cannot do

It looks exactly one turn ahead. It does not plan a sweep, scout, predict a double switch, save a Pokémon for later, or remember what worked last game. Its damage formula ignores items and every ability except Guts, and it knows nothing about the metagame beyond what that formula implies. Without the language model that the original puts at its core, it is a calculator with opinions about HP.

That is what we want from a reference opponent: simple enough to describe in five steps, so that a result against it means something.

Why it is in the rankings

It is a fixed yardstick. Ratings only mean something relative to opponents whose strength does not change. Our PokéChamp never updates, never gets deprecated and costs nothing to run, so every model added to the ladder, now or later, plays the same set of games against the same program. That is what makes a rating from one month comparable with a rating from the next.

It sits where the interesting comparisons are. Random AI marks the floor: a model that loses to it is broken. PokéChamp’s search marks something more useful, the strength of one-turn arithmetic done perfectly. A language model that beats it is doing something a calculator cannot. A language model that loses to it has read the whole battle, reasoned about it at length, and been outplayed by a handful of weighted numbers. The PokéChamp paper’s central point is that the best results come from combining the two.

It is rated like everyone else. Its rating is fitted from its games, with its own error bars, on the same scale as the models. It is not the fixed point of the scale: that is the anchor model, which sits at 1500 by definition. It has no rank number, because the ranking is for models; it appears on the leaderboard the way a reference line appears on a chart.

Statistics that do not apply

Our PokéChamp has never hallucinated a move, run out of tokens, timed out, or been rate-limited. It has also never explained a single decision. The autopilot, token and cost columns on the leaderboard are left blank for it, which we consider a fair summary of its personality.

The design, the name and the idea are PokéChamp's; the implementation, and any mistakes in it, are ours. Their own code is at github.com/sethkarten/pokechamp.