visnup/magical-athlete

Magical Athlete simulations

Raw output of a Monte Carlo simulator of the board game Magical Athlete (CMYK), following the August 2026 rulebook and official rulings: https://github.com/visnup/magical-athlete. Its README is the write-up these files back, and make there regenerates them.

Every file is one row per racer per race, or per player per game. Nothing is averaged.

The simulator, its analysis and this card were written with Claude Opus 5.5 at medium effort, working in Claude Code throughout.

Scoring a race

A game is four races: 1 and 3 on the Mild Mile, 2 and 4 on the Wild Wilds. Gold and silver chips are worth 3/2, 4/3, 4/3 and 5/4 in races 1 to 4, so a row's points in race k are gold * GOLD[k] + silver * SILVER[k] + bronze. The race files hold races on one track; score them as race 1 or 3 (Mild Mile) or 2 or 4 (Wild Wilds). For example, with DuckDB:

select racer, avg(gold * 3 + silver * 2 + bronze) points
from 'hf://datasets/visnup/magical-athlete/races-mild-5p.parquet'
group by racer order by points desc;

Reading the numbers

• Points per race, the average of that expression over a racer's rows, is the measure the write-up uses throughout, not win rate. Lovable Loser and Sisyphus top race 1 on bronze alone.
• A fair share is (GOLD[k] + SILVER[k]) / players, so 1.00 in race 1 at 5 players. Divide by it to compare table sizes.
• Noise. Each racer is in about 139,000 of a million 5-player races, so its points per race have a standard error of about 0.006 on the Mild Mile and up to 0.013 on the Wild Wilds (Scoocher in race 4). A pair of racers meets in about 16,000, so matchup differences under about 0.05 (Mild Mile) or 0.1 (Wild Wilds) are noise. The choices files run every option on the same seeds, good to about ±0.01. playouts is 4,000 races per racer, about ±0.1.
• Optional powers follow hand-written rules (choose in src/racers.ts), tuned against playouts. The racers that pick racers (Egg, Twin, Copycat, Mastermind) and the drafts use src/ratings.json, which comes from the previous run of the dataset, so it's one pass behind.
• Rules readings. Where the rulebook and rulings leave something open, the simulator's choice is listed under "Interpretations" in the repo's CLAUDE.md; sensitivity shows what flipping the open ones does.

Race rows

races-<track>-<players>p is a million races of a random lineup of 3 to 6 of the 36 racers. Seat 0 goes first. Egg, Twin, Mastermind and Copycat pick by the simulator's built-in ratings. Each race stands for either race on its track, so Twin gets the earlier winners it would have there: 0 or 2 on the Mild Mile (races 1 and 3), 1 or 3 on the Wild Wilds (races 2 and 4), drawn by each racer's chance of 1st and never from the race's own lineup. Columns:

column	meaning
race	race number within the file
seat	turn order, 0 first
racer	the racer
gold, silver	1 if it took that chip; a Mastermind that predicted itself can take both
bronze	bronze chip points gained (or lost) in the race
pos	final space, counting the Start as 0
trips	times it was tripped
uses	times its power was used
eliminated	1 if it was eliminated from the race
turns	turns it took
stalled	1 if the race ended without two finishers: a loop no one can finish (fiddly details 5), or the 1,000-turn limit

The other files with these columns each race one racer under test, or a rule variant:

• choices-<racer>: the racer under test in every race, with game_race (1 or 2) and choice, its optional power's decision rule (current is the simulator's default).
• playouts: the same, with playouts 0 for the hand-written choice or the number of playouts per option when it chose by playing each option out.
• twin: Twin, Egg, Mastermind or Copycat in every race of game_race 1 to 4, picking by ratings and, for Twin, with earlier winners to copy. twin-rated: whole races set up that way.
• lineup-twin: Twin in every race with earlier winners, used for its race-order ratings.
• sensitivity-<rule>-<track>: 5-player races with an unsettled rule read the other way (backwards: main moves pushed below 0 go backwards; queue: triggers queue and resolve even after the board changed).

Game rows

games-<players>p[-double], draft-5p and endgame-5p are whole games, a row per player:

column	meaning
game	game number
player	seat, 0 first
pick	first-round draft position, 1 first
team	its racers in the order raced, joined by |
estimate	the team's value by the simulator's ratings (games only)
score	total points
won	1, or a share of 1 when tied for the top score
strategy, me, met	draft and endgame: the test player's strategy, 1 on its row, and 1 if its M.O.U.T.H. raced an opponent's Scoocher

lineup-draft is computed, not simulated: each racer's points by race, its draft value (points a game over an average pick), and how often its best team plays it in each race.