Two layers, one truth

Underneath is a Bayesian skill engine (OpenSkill, the same family as TrueSkill). For each player it keeps two numbers: μ, its best estimate of your skill, and σ, how unsure it is. Every verified result updates both. Nothing else does.

The LR you see is a projection of that engine, not a separate score:

LR = 1000 + 40 × (μ − 3σ)

The −3σ matters: your displayed rating is the engine’s conservative estimate — a number it is confident you are at least. As you play, σ shrinks and LR rises toward μ even before you improve. That is placements working, without a single special-cased rule.

Because LR is a pure projection, the ladder can never disagree with the matchmaker — they read the same state. Platforms that keep a separate visible Elo eventually have the two drift, and then the ladder lies about who you’ll be matched with.

What moves it

Verified wins and losses. Nothing else. The size of each move depends on two things:

How certain the engine is

High σ (early games) moves you a lot; a settled rating moves slowly

Who you beat

Beating a favored team pays more; farming underdogs pays almost nothing

Those two together are why the ladder gets harder as you climb — not because we added a brake, but because at the top you are the favorite in most lobbies, so your average win is worth less and your losses cost more. The ceiling is earned against the odds, which is what makes it mean something.

What never moves it

These are deliberate exclusions with reasons, not features we haven’t built:

No performance points. Kills, damage, souls and MVPs never move LR. Per-stat rating invites stat-padding, mis-values roles — a peel-heavy support loses to a KDA farmer — and turns every teammate into a rival for rating. Performance gets its own surface (accolades and match MVPs), never rating weight.

No streak bonuses, no daily bonuses. Rating measures skill, not attendance or momentum. A win streak is displayed on your profile; it is never paid.

Rating is never a punishment. Bad behavior costs queue access through sanctions (the strike ladders), not points — deducting points would corrupt the skill estimate and the matchmaker with it. An abandon rates the match normally with the abandoner on their team’s result; the sanction is a separate track.

What a match is worth

You see the stakes before you play: the match card quotes the expected movement both ways (“win ≈ +22 · loss ≈ −26”), computed by running the real update in both directions before the game. Not a promise — a forecast from the actual engine.

And it is explained after: the match page shows what the result actually paid each placed player, with the why — an upset pays more than a favorite grinding out an expected win. Deltas are near-symmetric but not exactly, because the engine is updating its beliefs, not moving tokens between you.

A match is quoted only when it can be quoted honestly: if a side is entirely provisional, no figure is shown, because a number that lands nowhere near the actual change is worse than no number.

Placements

Placement matches

10 verified games

After a season reset

5 games to re-place

Starting point

Seeded from your in-game Deadlock rank, capped well below the leaderboard

While you place, your LR is hidden — from you and everyone else. It exists and it is moving; we won’t publish a number we don’t trust yet, because an early number is mostly noise and people anchor to it anyway.

What you get instead is a range — “tracking 1050–1400 LR” — the engine’s honest answer to “roughly where am I landing?”. The band is wide when σ is high and narrows every game, purely because the uncertainty burns off. When placements finish, the range becomes the number.

On match receipts the same honesty applies backwards: games you played while provisional show your placement progress, never a retroactive number — and the game that completes your placement shows where you arrived, not the hidden path there.

Levels

Ten levels sit over LR bands — identity, not mechanics. They gate nothing competitively; a hub may set a level floor for entry. The bands widen going up, and the top level is deliberately scarce.

Season 1 bands — recalibrated against the real distribution once the season has shape, and recalibration is announced, not silent.

Seasons and going quiet

Your rating never decays. Go quiet for 21 days and the engine slowly becomes less certain about you (σ inflates, capped), so your first games back move you faster — but the number itself was never confiscated while you slept.

The leaderboard has a separate visibility gate: at least one rated match in the trailing 14 days to appear. That kills ladder-camping without touching a single rating — the board is a picture of who is playing now; your rating is the record of how you play.

Seasons soft-reset: your rating is pulled partway toward the center and uncertainty is raised, then 5 matches re-place you. Season rewards key to your peak placed LR plus matches played — tanking at season’s end achieves nothing.

The ledger

Every rating change is an append-only ledger row: the match, the before, the after. Your profile graph is that ledger. Ratings can be replayed from it end to end — which is also how corrections work: if a match is voided after it rated, everything downstream is recomputed from the record, and the correction is appended rather than overwritten. The history shows what happened and what was fixed.

If this page and the platform ever disagree, that’s a bug worth reporting — the same standard the rules hold themselves to.