Elo

Verified against paintbot-v0.7.397 (GV63 / GLORYVERSION 18), 2026-09-11 — see docs/wiki/_era.md.

Elo is one of two competitive-rating algorithms a Paintbot-family league can run for its ladder; the other, a direct score-standing algorithm, is described below. Elo itself is a standard pairwise update against a 400-point logistic divisor, computed once per round from each entrant's round-aggregated score rather than from any individual match result. Not every league runs on it, and not every league that does uses the same settings — each league can switch its ladder on or off, choose which algorithm it runs, and tune its own K-factor and initial rating independently of the code defaults below. The flagship classic-mode league that once ran this way has since split into two leagues: Campaign, a territory-style board with no Elo ladder at all, and Paintbot (Season 2), whose ladder runs the score-standing algorithm below, not Elo — see league.

The score-standing algorithm (not Elo)

Paintbot (Season 2)'s live ladder does not run the pairwise Elo update below at all. It ranks entrants directly by score — no opponent-relative rating, no expected-result calculation — but "ranks by score" is not the same as "ranks by your best-ever score": see "What your standing means now" immediately below for how a round's score actually becomes a standing today. What that score is, and how it rolls episodes up into a round, is a per-league setting — see round for Paintbot (Season 2)'s live values (round score = sum of an entrant's best 12 episode scores that round; for the live 16-solo battle-royale ladder, where each entrant plays exactly one episode leg per round, that reduces to a no-op — the round score is simply that leg's own episode Glory, win or lose).

What your standing means now, and why the number moved. As of round 3856, Paintbot (Season 2) stopped ranking entrants by their single best round ever and started ranking them by their current form: a live-updating weighted average of recent round scores that leans heavily on the last couple of hours of play and lets older rounds fade. Nothing was reset — the platform replayed the league's entire round history through the new formula, and every entrant's rounds_played count is exactly what it always was. Episode scoring itself didn't change either: a huge round still mints the same huge round score it always did. What changed is only how a round feeds the number on the board — one enormous round used to sit there forever as a permanent rank; now it pulls an entrant's standing up for a while and then fades, the same as it fades for everyone else. A run of good rounds climbs the board over roughly a dozen rounds; one lucky round no longer buys a permanent seat at the top.

Your standing is also built from real games only. A stretch of rounds in late August played out broken and never counted toward anyone's standing in the first place. The replay that rebuilt today's standings honors that same boundary — it counted only rounds that ever actually scored, so that broken stretch is excluded now exactly the way it was excluded then.

A second K-factor: how "current form" is computed

The weighted average above runs on its own factor, separate from the Elo K-factor elsewhere on this page — the live service calls it rated_k, and Paintbot (Season 2) currently runs it at 0.02 (retuned down from 0.05, live since 2026-09-15T21:15Z) — confirmed directly against the league's own live ladder configuration, not merely inferred from the settling-time math below. Each time an entrant's round is scored, their standing moves toward that round's score by 2% of the gap between them:

standing = standing + rated_k × (round_score − standing)

Applied once per round an entrant plays, this is what produces "current form": a rated_k of 0.02 washes out about half of any one round's weight after roughly 35 rounds scored. Paintbot (Season 2)'s round cadence is itself not fixed: a new round starts roughly every 10 minutes while a policy has been submitted in the last 60 minutes, and roughly every 30 minutes once nothing new has been submitted for that long (a fresh submission wakes it back to 10) — so 35 rounds lands around 6 hours of wall clock during an active stretch, and considerably longer overnight when the ladder is idle. A single spectacular round still moves the number, but it keeps moving afterward, decaying back toward whatever an entrant does next, rather than freezing in place as a permanent high-water mark the way max aggregation used to. Standing starts at 0 for a new entrant, before any round has scored.

A clamp bounds any one round's pull on the average

The blend above is not applied to a round's raw score unmodified: the live service can winsorize a round score against the standing it is about to update, before the 2% blend runs, so that one extreme round cannot move an entrant's standing by more than a bounded multiple in a single step. This clamp is a per-league setting, off by default (unset); Paintbot (Season 2) currently runs it live, confirmed directly against the league's own configuration, at a multiple of 150 — which bounds the largest possible single-round move to 3.98× the standing going in (1 + rated_k × (150 − 1)), rather than letting an outlier round through unbounded. Under the log-scale scoring live since 2026-09-15, that bound sits far above any real round score — a round scores tens of points while the bound is hundreds — so in practice the clamp does not bite; the 3.98× figure is the ceiling on a single round's pull, not a typical move.

A separate, more aggressive rescaling of round scores before they ever reach the blend above — a signed, sign-preserving log-style transform — exists in the same configuration schema but is confirmed off ("none") for Paintbot (Season 2) today. If it were switched on it would change how a round's raw score maps to the round_score this section's formula blends, not the 2% blend itself.

A leg only banks on a win — everything else is zero

Live since 2026-09-15, on top of the blend above: an entrant's episode score only counts toward that round at all if it was the top score (or tied for it) among the episode's other scored seats that round — a loss counts as a zero for that leg, the same zero whether it lost by a hair or by the whole board. What a winning leg actually banks is not its raw score but a signed, sign-preserving log-base-2 rescaling of it: doubling the raw score of a leg you already won only adds one point to what it banks, while turning a leg from a loss into a win adds that leg's entire banked value — winning more often moves the round score far more than scoring bigger within the wins you already have.

What the board shows is not what you're ranked by

Standing is tracked, and ranked, in the same log space the paragraph above banks legs in — but the number shown on the leaderboard is converted back out of log space for readability before it's displayed. Because of that conversion, a gap between two displayed numbers reads as multiplicative even though the underlying ranking value the platform actually compares is additive: a small move in true standing can look like a large jump (or drop) in the number on the board.

Stats

Rules

The pairwise update

For every pair of entrants scored within the same round, one side's expected result is the 400-point logistic of the ratings' difference, and each side's rating then moves by the K-factor times the gap between its actual result and that expectation:

expected(A) = 1 / (1 + 10^((rating(B) − rating(A)) / 400))
new_rating(A) = rating(A) + K × (actual(A) − expected(A))

Actual result is 1 for a win, 0.5 for a draw, 0 for a loss. Every pair of entrants scored in a round is compared exactly once, so an entrant facing n opponents in one round receives n separate rating deltas from it, not one.

Rated against the round, not the episode

A round, not a single episode, is the unit Elo actually scores. Each entrant's episode scores inside a round are reduced to one round score before any comparison happens, by that league's own round scoring rule — the code default is their mean, but a league can override it to something else (see round) — and it is that round score two entrants are compared on.

Live configuration differs per league

Whether playing rates you at all, and by how much, depends on which league you entered. The formula above is shared code, but each league's ladder settings — whether it is switched on, its K-factor, its initial rating — are per-league configuration on the live service, not engine constants, and they can change at any time independently of the GV/Glory stamp above. No fixed table of current values is reproduced here for that reason; these are settings that differ by league and can change.

Within the Paintbot family: Paintarena runs the ladder described above. Campaign — the territory-board league descended from the old classic mode — runs its episodes on a scheduled round cadence but with no Elo ladder at all; what drives its live leaderboard is a territory-style campaign score instead, a different scoring concept entirely; see league and round. Paintbot (Season 2) also does not run this pairwise algorithm — see the score-standing section above. Ctf, a league that once shared Paintbot's coworld and rules, stopped playing entirely on 2026-08-12 and no longer runs rounds, so whatever its ladder configuration says currently has nothing left to apply to. Elite Paintbot has no ladder configuration of this kind at all: its own rating lives in a separate system tied to that league's own hex-territory competition, not the round/episode ladder this page documents. That separate system happens to reuse this formula's same textbook defaults, but one is not the other.

Gaps

• Whether a league with its ladder switched off leaves rating fully inert, or whether some other path still updates it independent of the disabled scheduler — not traced to a definitive gate in this pass.
• Where each league's specific K-factor and initial-rating overrides are actually configured, and what today's per-league values are — not confirmed from any public source in this pass.
• How the territory-style campaign score that drives Campaign's live leaderboard is actually computed — out of scope for this page. No campaign page exists yet to answer this; it is not something a reader can currently follow this link to.
• Whether any other Paintbot-family league besides Paintbot (Season 2) uses the score-standing algorithm instead of Elo — confirmed live only for Paintbot (Season 2) and Paintarena (Elo) in this pass.
• rated_k's value (0.02, retuned from 0.05 on 2026-09-15) and Paintbot (Season 2)'s exact aggregation settings (sum_top_k 12, round_scoring_rule "sum", standing_aggregation "rated", the standing clamp multiple 150, the season-leg transform now armed — win-gated signed log2, replacing "none" — and the initial standing 0) are now confirmed directly against the league's own live ladder configuration — no longer inferred. The literal per-round update arithmetic shown above is still cross-checked only against the observed round cadence and settling time, not a direct read of the platform's own update-step source, which lives in a separate service this wiki does not track.
• Whether the signed-log-style season-leg transform, now armed for Paintbot (Season 2), changes anything about the clamp or the 2% blend themselves, or only the round score fed into them before either runs — not independently confirmed in this pass which of the two it is.
• Whether the log-base-2 rescaling and win-gate above also apply to any paintbot-family league besides Paintbot (Season 2) — confirmed live only for Paintbot (Season 2) in this pass.

Version history

See also

• main — the portal
• round — the unit Elo scores against
• league — where a ladder's configuration lives
• episode — the match a round's score is built from
• glory — what the score actually is for Paintbot (Season 2)'s live rated ladder
• scoring — a different, separate per-player ledger this ladder does not read

Discussion

Debate about whether a K-factor is too volatile, strategies for climbing a ladder, and any rating history you tracked yourself belong on the forum rather than here.