Ranking

TinyBrains estimates each active version’s playing strength with TrueSkill. A ladder ranks versions by a conservative estimate, so an entry needs a high estimated mean and the matches to back it.

Reading a rating

A rating has a mean mu and uncertainty sigma. The displayed value is:

rating = mu − 3 × sigma

The initial prior is mu = 25 and sigma ≈ 8.333, which displays as a rating near zero. These are policy values. A hill score of 2 does not add 2 to your rating.

Each update depends on the match ranks, ties, the opponents’ estimates and the uncertainty. The hill-score margin plays no part. A surprise against a well-established opponent can carry different information from a result between two uncertain new versions.

One ladder, viewed by class

Every version has one rating, on Open, the single rated ladder. A match updates that one rating. A trial updates it for neither side.

A weight class is a view of Open: your class standing is your position on the Open ladder, counting only versions in your size class. So your Open rank and your class rank can never disagree about which of two same-size models is ahead. Compare models of different sizes on Open directly, since it already spans every size; a class view only shows how you rank among your size peers.

The leaderboard shows rank, model ID, owner, version, class, measured size, rating, provisional flag and match count. It sorts by conservative rating, highest first, and the version ID breaks ties the same way every time. Leaderboard rank and the shared game rank two seats get in a drawn match are separate numbers.

Why a result arrives before its rating change

The runner records the result first. Soma’s count clock then processes finished matches, writes rating events, and marks each match rated. Until then, the detail can show a result with no rating_change.

Each seat’s rating change records mu_before, sigma_before, mu_after and sigma_after for the one Open rating the match updated, so you can see how a result moved the estimate without comparing two leaderboard screenshots.

Provisional and settled

The provisional flag is true while sigma > 3, on your one Open sigma. The matchmaker also counts placement matches, and seeks at least eight before it settles a version.

In a season without rounds, a settled model can go without new matches of its own. The matchmaker can still pick it as an opponent, and those results move its rating. TrueSkill dynamics apply during updates only: uncertainty does not grow as wall-clock time passes. A season played in rounds works differently, as the next section explains.

Rounds and score resets

Without rounds, a version submitted in the first week collects hundreds of matches while one submitted later settles after a handful, and the conservative rating rewards the older one’s lower uncertainty. A season can be played in rounds to take a version’s age out of its score. The season’s rules say how often a round starts, typically weekly.

At the start of each round:

• every active version’s sigma rises to the round’s floor, so every version carries the same 3 × sigma discount and the leaderboard reads them by mu;
• a round may also draw every mu part of the way toward the season’s mean;
• matches queued for the round before are cancelled with ROUND_ENDED, while running ones finish and count;
• every active version then plays the round’s number of rated matches, the least-played first.

Your match count, your record and your rating history carry over; only the uncertainty and possibly the mean are reset. The leaderboard shows each version’s matches in the current round beside its season total, and counts down to the next reset. A notice across the site, and one in your notifications, gives warning before each reset.

Finals

When a season’s submissions close, an administrator can start its finals, a last round with its own number of matches. The finals wait until every submission has been admitted, so every entry’s last version plays them. They start from a reset like any round’s. Then every entry plays exactly the finals’ number of matches: no entry is seated past it, while baselines fill the seats and are never waited for. The season closes once every entry has played its matches and nothing is left in flight. The podium is frozen from the ratings the finals end on.

An administrator can change the finals’ number of matches while they run, for every entry at once, or end the season early.

Ratings after a new version

A successor that passes its trial inherits its predecessor’s one Open rating mean in the same season. The predecessor is the same model’s previous active version; your models are separate lineages and inherit nothing from each other. The platform doubles the successor’s uncertainty, capped at the initial prior, so it needs new evidence, and its displayed rating can start lower than the predecessor’s with the same mean.

The successor inherits that one rating whatever weight class it lands in. Each version keeps its own later results: the platform never moves a predecessor’s in-flight match to its successor.

Seasons and comparisons

Read one season with GET /v1/games/ants/leaderboard?ladder=open&season=<slug>. Without season, the API answers for the game’s featured season. A private season’s standings are its members’ alone, through /v1/private/games/ants/leaderboard. Historical standings stay available, and after an administrator closes a season, matches already in flight can still add final updates. A rating compares you with that season’s field; ratings from two seasons share no scale.