Output above a replaceable player.

Every league has a freely available floor: the players a team could sign tomorrow with no real cost. Call that replacement level, concretely, the 25th percentile of qualifying players at each position. CAR is how far a player's output rises above that floor, a composite score on an artificial scale, not a count of wins. A replacement-level player scores 0. The rest is what you actually bought.

It isolates the player from the team.

Raw stats reward players on winning teams and punish players on losing ones. So CAR is built almost entirely from numbers that cannot carry team level: a player's share of his own team's farm, gold or kills; his damage per unit of gold spent; his share of the damage against his share of the dying; and his lane result measured against the opponent standing in front of him. A share cannot rise just because the team is winning, and a duel with the enemy laner is a duel whoever else is on the map.

Where a share is shaped by circumstance, CAR scores it against expectation instead of deleting it. A top laner on a team that barely touches towers will post a huge slice of a tiny pie; one on a dominant team is buried in a big denominator. So tower pressure is credited as output above what an average player would post with those same teammates — the environment becomes the baseline, not the signal. And a stat only earns a place in a formula if it moves with a player's own good and bad stretches: habits that merely differ from player to player — however reliably — describe style, and stay out of the score.

Every role is built separately, because the same statistic means different things in different seats. Gold difference at ten minutes is close to a clean duel result in the top lane and nearly meaningless in the bot lane, where two players and a jungler decide it.

The same logic is applied to the champion itself. Some champions inflate the box score through their kit alone, not through better play, so CAR treats every champion as a park factor and strips that inflation out before anything else is computed. What remains is a champion-neutral estimate of the player, and it is worth seeing how that is done.

How the champion is taken out.

A support on Nami and a support on Pyke are not playing the same statistical game. One of them places wards because that is the job the kit hands him; the other roams and gets credited for it. Ivern will never post damage numbers, Zyra cannot help posting them. Senna gives up early gold by design. None of that is a judgement about the player, and all of it lands in the box score, so it is measured and removed before a single component of CAR is computed.

The estimate is made on the champions, not on the players. For every raw statistic the formulas use — damage per minute, gold and damage share, kill participation, deaths, wards, gold and CS difference in lane, tower damage — the model fits a per-champion effect from every game in the tracked leagues: how much does this champion move this number, on its own? Each effect is fitted inside its own role, region and season, because a champion is a different tool in the top lane than in the bot lane, and a different tool in one patch cycle than the next. Thin cells are pulled toward zero: a champion picked six times does not earn a strong correction, it earns a cautious one. Effects are refitted each split and carried forward with the neighbouring splits, so a champion that drifts with the patch drifts in the model too.

Then it is subtracted, player by player, through the pool he actually played. A player's correction is the average of his champions' effects across his own games, weighted exactly the way the statistic itself is built — by minutes for a per-minute rate, by his team's kills for a kill share, by games for a lane differential. Play ten games of the champion and you carry ten games of its effect; play it once, and once. That subtraction happens on the raw season totals, before damage efficiency, damage-per-death, kill participation over expectation and the vision terms are derived, so every downstream number is built on champion-neutral inputs rather than patched up afterwards.

The honest limits. A champion's effect is fitted from everyone who played it, which includes the player being corrected, so a player who is effectively the only one on a champion is partly measured against himself. In practice that is rare, and it concentrates in tiny samples: across a league season, about one game in a hundred sits in a champion cell that a single player owns outright. And neutralization deliberately does not credit the pick. Choosing the champion that inflates your numbers is a real skill; it is not the skill CAR is trying to measure.

Don't compare across regions.

Replacement level is recomputed inside each league and split, and cross-region samples are reset. A +30 in the LCK and a +30 in a development league are not the same achievement; they're measured against different floors. CAR ranks players within a competitive environment, never across them.