Ranking

TinyBrains uses TrueSkill to estimate each active version’s playing strength. A ladder ranks versions by a conservative estimate, so an entry must provide evidence through matches as well as obtain a high estimated mean.

Reading a rating

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

rating = mu − 3 × sigma

The current initial prior is mu = 25 and sigma ≈ 8.333, producing a displayed rating near zero. These are policy values, not game points. A hill score of 2 does not add 2 to your rating.

Match ranks, ties, opponent estimates, and uncertainty determine each update. The margin of the hill score is not an additional input. A surprising result against a well-established opponent can carry different information from a result between two uncertain new versions.

Your ladders

Every version has its size-class ladder and Open. All-same-class matches update both; mixed-class matches update only Open. Trials update neither ladder, even for the opponent. Compare different-sized models on Open rather than comparing numbers from unrelated class ladders.

The leaderboard shows rank, model ID, owner, version, class, measured size, rating, provisional flag, and match count. Its order is descending conservative rating, with model ID as a deterministic tiebreaker. This leaderboard position differs from shared game ranks inside a drawn match.

Why a result arrives before its rating change

The match worker records a result first. A separate counting step processes finished matches and writes rating events, then marks the match rated. Until then, detail may show a result with no rating_change.

Per-seat rating changes record mu_before, sigma_before, mu_after, and sigma_after for each ladder updated. They explain how a result changed the estimate without requiring you to infer it from two leaderboard screenshots.

Provisional and settled

The current provisional flag is true when sigma > 3. Scheduling also considers placement match counts: current policy seeks at least eight matches on ladders the version can actually reach. A model alone in its class is evaluated for settling on Open until another active same-class version exists.

A settled model need not receive a continuous stream of new matches. It can still play as an opponent, and its rating can change with those results. The current implementation applies TrueSkill dynamics during updates; it does not inflate uncertainty merely because wall-clock time passes.

Ratings after a new version

A successful successor inherits the predecessor’s mean on shared ladders within the same season, where the predecessor is the same model’s previous active version. Your models are separate lineages and inherit nothing from each other. Its uncertainty is doubled, capped at the initial prior, so new evidence is required. The initial displayed value can therefore fall even though the inherited mean stays the same.

If the successor changes weight class, the new class rating starts from the prior; Open can inherit. Each version’s subsequent results belong to that version. A predecessor match already in flight is not reassigned to its successor.

Seasons and comparisons

Use GET /v1/games/ants/leaderboard?ladder=open&season=<slug> for a specific season. Without season, the API chooses the live season, or the latest closed one if none is live. Historical standings remain available; after an administrative close, already in-flight matches may still contribute final updates. Treat ratings as comparisons within that season’s field rather than a universal scale across seasons.