Carry Score Methodology

Most post-game ratings collapse League into the same few numbers: KDA, damage, CS, and win or loss. That creates bad incentives. A Janna should not be punished for having fewer kills than Zed, and Zed should not get credit for ignoring side pressure because he padded damage into a tank.

Carry Score is OneLifeHero's attempt to describe win impact in the language of the champion being played. It asks what that champion was supposed to do in that role and build, then weighs the match around those expectations.

The score is built from six visible pillars: Early/Lane, Combat, Objective, Map Pressure, Utility/Control, and Discipline/Efficiency. They are not all equally important for every champion. A jungle carry can get more credit for tempo and objective conversion, while an enchanter can earn more from vision, survival, shielding, and fight control.

The UI intentionally avoids showing fourteen categories at once. A score is only useful if the player can understand the story quickly: what carried, what failed, and what made this game different from the last one.

The model detects archetypes from champion identity, role, items, and runes when the data exists. AP Shaco, AD Shaco, enchanter Seraphine, AP Seraphine, tank Gragas, and AP Gragas should not be judged as if they were the same job with different cosmetics.

That also means losses can still score highly. A player can hard-carry a losing game if they created enough pressure, won enough fights, or kept the map alive while teammates collapsed. The opposite is also true: a player can win as a passenger and receive a modest score if the team did the real work.

Each scored game gets a style title, top reasons, and a weakness. The system should not force positivity. Some games deserve three strong reasons; some games deserve one dry note and a clear weakness. The goal is not to flatter the player. The goal is to make the card feel like a sharp analyst watched the match and understood the champion.

Future calibration can improve the score with percentiles and larger benchmark pools. The current version stays deterministic and explainable: it uses match data and timeline data where available, lowers confidence when timeline data is missing, and stores computed rows so later versions can be compared instead of silently changing history.