eukleos
Method
Every rating on eukleos comes from the full finishing order of real pro lobbies. Every claim about its accuracy is measured on games the model had not seen, and published with its uncertainty.
What is rated
A TFT lobby is eight players finishing first to eighth. eukleos reads the whole order, not just top four or points, with a Plackett-Luce model: every player has a strength, first place goes to one of the eight in proportion to their strengths, second place to one of the remaining seven, and so on down to eighth.
Strengths drift slowly over time, so recent games count more than old ones. Every player has one continuous rating across sets.
The scale
A rating is 1500 plus 1000 times the player's log-strength. A 100-point gap is a small edge: about 53 versus 50 percent to finish top four against an average field.
The ± next to a rating is its 95% range. Two players whose ranges overlap cannot yet be told apart. Ratings with fewer than 20 games behind them are provisional and shown in grey.
How well it predicts
Each event in Sets 16 and 17 was predicted using only games played before it, then scored against what happened: 4,136 lobby-games in all. The score is how much more likely the model found each lobby's real finishing order than a random guess, in thousandths of a nat per lobby.
eukleos scores highest, but its lead over OpenSkill, Elo and average placement is within the noise. It is chosen for modelling the whole lobby and for giving every rating an honest range, not for a proven accuracy win. Any single TFT game carries a lot of luck.
How many games a rating needs
Reliability is the share of a rating's spread that reflects real skill rather than luck. It grows slowly with games played.
Even past 160 games, about half of a rating's spread is still luck. Read close rankings as ties.
One rating across sets
Resetting every player at the start of a set loses about nine tenths of the rating's predictive edge. A new set is not measurably more disruptive than the steady drift within a set, so eukleos carries ratings over and lets them adjust.
Predictions made in advance
Predictions for the Set 18 Tactician's Crown were frozen and committed on November 8, 2026 (commit bcf7d51), before the event, together with Elo and OpenSkill predictions and the code that will score all three.
The favourite, Dishsoap, is given a 5.6% chance to win the Crown, against 2.5% for a player picked at random. In an eight-player game, even the best player usually does not win.
What eukleos does not claim
Strict rank orders between close players, that the model beats Elo or OpenSkill, or that one region is stronger than another. Only 115 lobby-games in the test mixed regions, too few to compare them.
Data
Lobby results come from Liquipedia (CC-BY-SA 3.0) and Riot's official standings sheets, checked against each other. eukleos does not use the Riot API or ranked ladder games.
Current model pl_combined@6134a68831e7, data through 2026-11-07: 10,123 lobby-games, 2,543 players, 200 events.
Decision report, every formula in plain language and the source code.