Evidence has to earn a recommendation.

Two decisions, two models

Runes are chosen from information legitimately available at champion-select completion: champion, role, lane opponent when identifiable, both team compositions, queue, patch, and a broad skill band. Items are a different problem. They are ranked at a shop decision from the observable game state up to that moment, including inventory, components, gold, levels, prior combat, objective state, and damage profile.

Itemized does not treat the most common winning build as automatically optimal. Winning players often buy differently because they are already ahead; careless analysis mistakes that effect for item strength.

What the system produces

• A clearly ranked primary rune setup before the game, with supported alternatives when the tradeoff is meaningful.
• A primary next purchase for the observable state at a shop decision, plus viable conditional alternatives.
• Internal support, uncertainty, patch-recency, and rank-coverage checks before an answer is shown.
• An explicit “insufficient evidence” result when no candidate is supported reliably.

Itemized is decision support, not a decision replacement. The interface makes the default answer fast to read while preserving player agency by exposing alternatives and the condition under which each becomes preferable. The player always makes the final rune and purchase choice.

Statistics are evidence, not instructions

The champion workspace reports descriptive win rate, pick rate, ban rate, games, rune pages, matchups, purchase positions, and exact item sequences for the selected population. These tables remain separate from recommendation scores. A high observed win rate does not by itself establish that a choice caused more wins.

Item tables preserve purchase order and sample size. Later-item results are explicitly conditional on players reaching that purchase, so their win rates cannot be compared directly with first-item results. The prototype table also shows deviation from the 50% match baseline. Production rank and region filters will use independently measured cohorts rather than recalculating one snapshot cosmetically.

Leakage boundary

A feature is allowed only if it existed before the decision. Match outcome and post-decision statistics are labels or evaluation data, never model inputs. Item events use the previous timeline frame, accepting some staleness instead of leaking future information. End-of-game inventory is reserved for auditing.

Evaluation gates

1. Popularity and empirical-Bayes win-rate baselines establish a minimum bar.
2. Outcome and action-propensity models adjust for observed state and measure action support.
3. Doubly robust off-policy evaluation estimates policy value with uncertainty and overlap diagnostics.
4. Chronological holdouts test later patches and prevent the same match or player neighborhood leaking across splits.
5. A frozen shadow test compares identical game states against established public recommendation tools.
6. No superiority claim is made without a prospective, randomized evaluation.

Data sources and integrity

The research backend uses Riot's supported Account-v1, League-v4, Match-v5, Summoner-v4, and Data Dragon services. The public product will not expose hidden information, scout an opponent's current session, read game memory, estimate private MMR, automate a purchase, or control gameplay. Recommendations are derived from aggregate historical data and context visible to the player. The player chooses whether to follow each recommendation.