PaperlayerView exact revision on arXiv

Source-linked AI summary

When the Rule-Maker Runs the World Championship: Late Patches and Procedural Accountability in League of Legends

Haewoon Kwak

arXiv:2610.07427v1cs.CY

TL;DR

Because Riot both changes League of Legends shortly before Worlds and runs the championship, the paper asks who scrutinizes how late changes distribute burdens. It measures team exposure using team-specific patch windows and prior pick shares, finding more negative net exposure among teams reliant on contested champions and an association between buff exposure and additional opponent bans.

  • Problem

    Because Riot both changes the game shortly before Worlds and runs the championship, who scrutinizes how late changes distribute benefits and burdens?

    Source: §1 · p. 1, §5 · p. 12

  • Method

    The study measures team exposure using team-specific patch windows and prior pick shares, and examines how pool breadth and meta concentration relate to net exposure.

    Source: §1 · p. 2, §4.1 · p. 9

  • Results

    Teams reliant on widely contested champions had more negative net exposure, while greater buff exposure was associated with additional opponent bans on buffed champions.

    Source: §1 · p. 2, §5 · p. 12

  • Takeaways & Limitations

    Scrutiny should begin before play through earlier commitment, explicit team-level incidence assessment, independent scrutiny, and a usable process for challenging decisions.

    Source: §7 · p. 15

    Only ten events support the event-bootstrap intervals, teams recur across them, and uncertainty, model misspecification, and observational confounding can affect inference.

    Source: §6 · p. 14

Abstract

from arXiv · show

Riot Games, which runs the League of Legends World Championship (Worlds), also changes the game shortly before each tournament. This concentration of authority raises a governance question: Who scrutinizes how the benefits and burdens of late rule changes are distributed? We examine all ten championships from 2016 to 2025, combining 1,097 games, 215 team-years, and LLM coding of 1,090 champion changes across 40 patches. On average, 33% of a team's prior picks were on champions with net nerfs in the patches it had not played before Worlds, and 8% were on champions with net buffs. Within a year, the median gap between the most and least affected teams was 23 percentage points for nerfed picks and 12 for buffed picks. Nerfs fell mainly on the most-contested champions and buffs on less-contested ones, as Riot's published balancing criteria anticipate. Pools concentrated on the prevailing meta therefore had more negative net exposure, and even the winners of a year's four major regional leagues differed in net exposure about as much as four randomly drawn teams. Higher buff exposure was associated with additional opponent bans on buffed champions, although its size depends on the model. Among quarterfinalists, the most buff-exposed team of each year reached six of ten finals and won four titles. The team with the most negative net exposure reached two finals and won one. These exploratory descriptions identify unequal exposure, not causal gains or losses. The association between net exposure and winning remains uncertain, at +1.3 percentage points per standard deviation with a 95% interval from -0.9 to +3.7. We argue that scrutiny of late rulemaking should not depend on proving that a patch changed who won Worlds. Earlier patch commitments, team-level impact disclosure, independent review, and a formal opportunity to challenge decisions would make publisher authority more accountable.

1 Introduction

Riot both changes League of Legends shortly before Worlds and runs the championship, concentrating authority over competitive conditions. This study examines how late changes intersect with team pools and argues for accountability before play, without requiring proof that a patch changed the winner.

1 Introduction

  • Riot’s published patch notes and balancing criteria explain changes and goals, but do not show how objectives were weighed, which teams bear costs, or how participants can seek independent review.Common patches can affect teams differently because they qualified with different pools and often played different game versions across regions.Source: §1 · p. 2
  • The study examines all ten Worlds from 2016 to 2025, measuring exposure from prior official picks and team-specific patches not played before Worlds.The measure does not capture lost champion mastery, adaptation practice time, or changes in title probability.Source: §1 · p. 2
  • The authors argue that late championship rulemaking should bring obligations to disclose team-level incidence, justify tradeoffs, and provide a meaningful route to challenge decisions.Source: §1 · p. 2
  • Teams reliant on widely contested champions had more negative net exposure, while the winners of regional leagues differed about as much as randomly drawn teams.Nerfs concentrated on the most-contested champions and buffs on less-contested ones, consistent with Riot’s published criteria.Source: §1 · p. 2
  • Greater buff exposure was associated with additional opponent bans on buffed champions, but the association’s size depended on how pick share was modeled.Source: §1 · p. 2
  • Match-result associations remain uncertain, so the outcome analysis does not settle disputes about late patches in either direction.Each title is a single event without an observable counterfactual.Source: §1 · p. 2

2 Background and related work

League of Legends teams compete through champion drafts, while Riot changes champion statistics and abilities through patches. The background places this setting alongside professional-play balancing criteria, research on patch effects, sports rule changes, and esports governance.

2 Background and related work

  • Teams draft by banning and picking champions; patches strengthen champions through buffs or weaken them through nerfs.A picked or banned champion is unavailable to the other team in that game; in 2025, fearless draft also made champions picked earlier in a series unavailable.Source: §2 · p. 3
  • Riot’s professional-play criteria use aggregate champion presence and performance to guide nerfs and buffs, but do not account for how changes fall on teams with particular pools.Worlds uses a designated patch, while participants’ last officially played patches can differ.Source: §2 · p. 3
  • Prior League of Legends studies examine patch-related champion win rates and performance variation; this paper measures team-specific exposure and distinguishes it from draft associations and outcome uncertainty.Source: §2 · p. 3
  • Sports-economics research treats advantage distribution and outcome uncertainty as central to contests, and studies how rule changes can alter incentives and competitive balance.Worlds combines frequent software revision, a short pre-championship interval, and publisher control of both game design and event operation.Source: §2 · p. 3
  • Esports governance scholarship emphasizes publisher ownership of the intellectual property underlying professional competition; this paper connects that concern to the distribution of pre-event changes across teams.The authors do not measure audience trust, participant perceptions, or the effectiveness of existing consultation, and do not assume esports lacks all independent dispute resolution.Source: §2 · p. 3

3 Data and measures

This section describes the match data, patch coding, exposure measures, and analytical methods used to study Worlds. It combines restored team histories with blinded coding and specifies how exposure, pre-tournament strength, and draft and outcome associations are measured.

3.1 Matches

  • The final dataset contains 1,097 games and 215 team-years, including 547 restored pre-Worlds games; matched sources agreed on winners in 99.99% and bans in 99.7%.All team-years had at least ten official games in the preceding 100 days, so all 1,097 games enter the main exposure–outcome model.Source: §3.1 · p. 4

3.2 Patch window and blind coding

  • Across 40 patches, 1,090 Summoner’s Rift champion changes were coded from patch notes, with agents instructed not to access esports results, winners, or professional picks.Exposure is restricted to direct champion changes; omitted item, rune, and system changes could reinforce or offset them.Source: §3.2 · pp. 4–5
  • Direction agreed in 87.5% of author-coded changes and sign in 89.2%; repeat coding agreed on sign in 99.4%, while coded changes tracked presence shifts.Buffed champions rose 7.8 pp in pick-and-ban presence and nerfed champions fell 12.6 pp; main-result directions and statistical conclusions survived sensitivity recodings.Source: §3.2 · pp. 4–5

3.3 Exposure

  • Exposure weights prior 100-day, role-specific pick shares by changes on unplayed patch versions; separate buff and nerf scores and unweighted pick percentages complement standardized scores.Teams’ windows covered zero to six coded patches (mean 3.2).Source: §3.3 · pp. 5–6

3.4 Pre-tournament strength

  • Pre-tournament strength comes from a Bradley–Terry model fitted to each team’s prior-year games; it classified 68.7% of Worlds games correctly and underrated LCK teams by 5.2 pp per cross-region game.Aggregated forecasts classified 72.3% of series correctly.Source: §3.4 · p. 6

3.5 Inference

  • Game-level logistic models estimate exposure differences in win probability, with 95% intervals resampling whole events; studentized within-year permutations do not guarantee validity for observational exposures.Secondary models are exploratory, p-values are not corrected for multiple analyses, and post-hoc tournament finishes are descriptive.Source: §3.5 · pp. 6–7
  • The ban analysis counts opponents’ bans on champions picked in at least 10% of a team’s recent games, subtracting expected bans based on other games at that Worlds.Champions unavailable for selection are excluded from both counts.Source: §3.5 · p. 6
  • Pool breadth is the average effective champion count across roles, while meta concentration weights champions by their pick-and-ban presence among other teams.The effective count is the number of equally played champions that would yield the same entropy.Source: §3.5 · pp. 6–7

4 Results

Across ten championships, late patches affected teams’ established champion pools unevenly, with exposure shaped by pool style and associated with opponent bans. Match results do not resolve whether exposure affected who won, while prominent team cases remain descriptive.

4.1 Who pays for balance? Nerfs follow the contested meta

  • 33% of pre-Worlds picks were on net-nerfed champions and 8% on net-buffed champions; median within-year team gaps were 23 and 12 percentage points, respectively.Exposure also varied by role: nerfed champions accounted for 46% of bottom-lane and 41% of jungle picks, compared with 24% in top lane and 22% in support.Source: §4.1 · p. 7, Figure 1 · p. 8

4.2 Do teams get to use their buffs? Bans follow buff exposure

  • Top-third buff-exposure teams faced 0.22 opponent bans per game versus 0.10 expected, an excess of 0.12; bottom-third excess was approximately zero.The estimated association between buff exposure and bans depends on how pick share is modeled.Source: Figure 1 · p. 8, §4.2 · p. 10

4.3 Does exposure change who wins? The evidence is inconclusive

  • Net exposure was associated with +1.3 pp win probability per SD (95% interval −0.9 to +3.7), leaving the association with winning unresolved.Draft measures also had intervals including zero, and the games would usually detect only associations around 3.4 pp per SD or larger.Source: Figure 1 · p. 8, §4.3 · pp. 10–11

4.1 Who pays for balance? Nerfs follow the contested meta

  • In the most-contested decile, 52% of champions were nerfed and 1% buffed, versus 6% nerfed and 23% buffed in the bottom eight deciles.The pattern accords with Riot’s published presence-based criteria, though those criteria do not specify how changes affect teams’ different pools.Source: Figure 2 · p. 9, §4.1 · p. 9
  • A one-SD increase in meta concentration predicts +0.35 SD nerf exposure and −0.40 SD net exposure; its buff-exposure association was not robust across presence definitions.Pool breadth added no detectable association after meta concentration was included.Source: §4.1 · p. 9
  • Pool breadth and meta concentration ranked held-out teams’ net exposure with mean Spearman correlation 0.35 (range −0.31 to 0.61), correctly ordering nine years.The model was fitted on the other nine years; 2020 was the exception.Source: §4.1 · p. 9

4.4 Which teams stand out? The most buff-exposed and most negatively exposed quarterfinalists

  • Among quarterfinalists, the yearly most buff-exposed teams reached six finals and won four titles; the most negatively exposed team won once.These are descriptive counts, not tests against a benchmark.Source: §4.4 · pp. 11–12
  • Across all teams, yearly most buff-exposed teams reached two finals and won one title; lowest-net-exposure teams exited before quarterfinals five times and never won.The two exposure perspectives support scrutiny across the full field and both directions of change.Source: §4.4 · pp. 11–12

5 Discussion

Late patches make authority over competitive conditions a governance question because they intersect unevenly with team pools. The evidence supports procedural safeguards before play, without requiring proof that a patch changed who won Worlds.

5 Discussion

  • Late rulemaking is governance because changing champions shortly before Worlds changes the resources teams bring to competition.The proposed scrutiny concerns the decision process, not proof that a patch changed a title result.Source: §5 · pp. 12–13
  • Teams reliant on widely contested champions had more negative net exposure, while higher buff exposure was associated with additional opponent bans on buffed champions.The size of the ban association depends on how pick share is modeled; legitimate balance goals can still impose unequal adjustment demands.Source: §5 · pp. 12–13
  • The analysis cannot resolve whether exposure changed who won: even across ten events, the association between exposure and winning remains uncertain.A single championship has no observable counterfactual, and the illustrative power analysis detects smaller associations less than 80% of the time.Source: §5 · p. 13
  • A governance record should explain the intended problem, alternatives, anticipated team-level incidence, and reasons for making changes close to Worlds.After teams’ last official games, exposure can be computed exactly before the tournament; disclosure should distinguish buffs from nerfs and acknowledge omitted item and system changes.Source: §5 · p. 13
  • The proposed safeguards pair an earlier patch timetable and defined freeze with impact statements, independent review, and a time-bounded route to challenge decisions.Exceptions for critical bugs or competitive exploits should have published reasons; the analysis does not identify an optimal freeze duration.Source: §5 · pp. 13–14
  • Review should use common criteria, not give adversely exposed teams a veto or make equal numerical exposure the objective.Earlier freezes can preserve imbalances that later patches might correct, while teams have incentives to defend preferred champions.Source: §5 · p. 14

6 Limitations

The paper’s inferences are constrained by a small number of events, measurement uncertainty, and limits in the match and patch data. The authors also bound what their analysis can establish about publisher conduct and the proposed safeguards.

6 Limitations

  • Only ten events support event-bootstrap intervals, teams recur across them, and inference is also exposed to measurement uncertainty, dependence, misspecification, and confounding.The conditional power simulation does not propagate uncertainty in ratings or exposure; some comparisons are exploratory and based on quarterfinalists selected after the fact.Source: §6 · p. 14
  • Exposure counts direct champion changes, not item, rune, or system changes that may reinforce or offset them.One author coded changes, magnitude codes are less reliable than directions, and pre-Worlds pools may already reflect adaptation to announced patches.Source: §6 · p. 14
  • The meta-concentration association with buff exposure depends on presence definitions and within-Worlds dependence, unlike its associations with nerf and net exposure.The size of the ban association also depends on how pick share is modeled.Source: §6 · p. 15
  • The restored 547 games cannot all be checked against Oracle’s Elixir, and disagreements between sources lead to excluding games that were played.The sources are not fully independent, and games missing from both sources are not recovered.Source: §6 · p. 15
  • The study does not assess whether individual patches followed the cited 2020 criteria or establish publisher intent, causal title effects, perceived fairness, or safeguard effectiveness.Source: §6 · p. 15

7 Conclusion

Across ten Worlds, late changes intersected unevenly with team pools and were associated with differences in opponent bans. The match-results analysis does not resolve whether this distribution changed who won, so the authors argue for procedural accountability before play.

7 Conclusion

  • Late changes intersected unevenly with team pools, and net exposure was more negative for teams concentrated on the contested meta, including regional winners.Source: §7 · p. 15
  • Higher buff exposure was associated with differences in opponent bans, but match results do not resolve whether exposure changed who won.Source: §7 · p. 15
  • The authors propose earlier commitments, explicit assessment of team-level incidence, independent scrutiny, and a usable challenge process before competition.The stated aim is to make the exercise of publisher authority visible, justified, and contestable.Source: §7 · p. 15

Ethics statement

The study uses public match records and patch notes, with no private data or interaction with players, teams, or the publisher. Naming teams is intended to support scrutiny of specific championships, not to question their results.

Ethics statement

  • The analysis uses publicly available professional-match records and released patch notes, with no private data or interaction with participants or the publisher.Source: Ethics statement · p. 15
  • Teams are named because the governance questions concern specific public championships, not to diminish their achievements.Source: Ethics statement · p. 15
  • The authors report no evidence that any title was undeserved or that the publisher favored a team; T1 is included as an illustrative high-buff-exposure case.Source: Ethics statement · p. 15

Data and code availability

The paper’s match data and derived materials are publicly available through the cited sources and a repository.

Data and code availability

  • Match data come from Oracle’s Elixir, with restored games from Leaguepedia included under its CC BY-SA 3.0 license.The restored games are included in the repository.Source: Data and code availability · p. 15
  • The repository makes the coded patch notes, validation data, analysis code, and derived tables available.Source: Data and code availability · p. 16

Use of AI tools

The author describes AI assistance in coding and writing, alongside checks and documentation for the analyses. The appendix details restoration and correction of pre-Worlds match data.

Use of AI tools

  • The author says AI assistants helped with coding and writing, while the author checked all code, analyses, and text and takes responsibility.LLM agents also coded patch notes, as described elsewhere in the paper.Source: Use of AI tools · p. 16

A Games restored from Leaguepedia

  • The restoration process adds 547 participant games, chiefly from earlier years, and changes the last played version or 100-day pools for some team-years.It changes the last officially played version for seven team-years and the 100-day pools for 55 team-years.Source: A Games restored from Leaguepedia · p. 19
  • Oracle’s Elixir records were checked for copies and winner or team-label problems, with rules applied to remove, set, exclude, or move records.Two participant games remain missing, including one absent from both sources’ scoreboards.Source: A Games restored from Leaguepedia · pp. 19–20

B Patch-note coding

LLM agents coded champion changes from patch notes using a defined direction and magnitude scheme, with audits and independent checks. Alternative codings preserved several main patterns, while the buff coefficient proved coding-sensitive.

B Patch-note coding

  • Across 40 patches, agents coded each changed champion’s net direction as buff, nerf, adjust, rework, or bugfix, assigning magnitude levels to buffs and nerfs; non-Summoner’s Rift changes were excluded.Source: B Patch-note coding · p. 21, Table 4 · p. 22
  • The author’s stratified sample agreed with LLM coding on sign in 89.2% of 120 changes, while a separate same-model coder agreed on 99.4% of 309 pairs.The audit found no esports-results site access, although blindness was procedural rather than absolute because models may carry prior knowledge.Source: B Patch-note coding · pp. 21–22
  • For 434 one-directional changes coded as buffs or nerfs, coded signs matched game-data directions in 98.4%.This check covered comparable game-data changes and did not capture changes absent from those data.Source: B Patch-note coding · pp. 22–23, Table 5 · p. 23
  • Under coding variant B, the separate-exposure buff coefficient fell from +2.2 pp under A to +0.7 pp, so the coefficient is sensitive to coding.Variant B drops or reverses many buffs that game data do not show.Source: B Patch-note coding · p. 23, Table 6 · p. 24
  • Across three alternative codings, the nerf concentration, buffed-champion ban interaction, outcome interval including zero, and roughly 3.4 pp/game detection threshold remained.The most buff-exposed quarterfinalists still reached six finals and won four titles; the most negatively exposed reached two finals and won one under two variants.Source: B Patch-note coding · p. 23, Table 6 · p. 24

C Pre-tournament strength and robustness

The robustness analyses test alternative controls, presence references, team-lineage definitions, and role-specific exposure. Nerf and net associations are more stable across checks than buff associations, while organizational differences remain uncertain.

C Pre-tournament strength and robustness

D Presence reference and dependence within a Worlds

  • Nerf and net associations hold across presence-reference checks, while the buff association weakens under Worlds-level inference and disappears when presence is measured before the first change.The two-way and bootstrap p-values for buff exposure are 0.048 and 0.096, respectively.Source: D Presence reference and dependence… · pp. 25–26, Table 8 · p. 26

E Team lineages and persistence

  • Buff exposure persists less clearly within linked team lineages than breadth or meta concentration, and evidence of organization-specific differences depends on lineage definitions.The linked-organization result rests largely on one lineage; the test establishes neither organizational differences nor their absence and cannot test intent.Source: E Team lineages and persistence · pp. 26–27, Table 9 · p. 26

F Role-level exposure

  • Nerfed-pick shares are highest in bottom lane and jungle and lowest in top lane and support, while buffed-pick shares are highest in top lane.Among team-years with nonzero exposure, the largest-exposure role accounted for a median 40% of the sum of absolute role exposures, versus 20% under even distribution.Source: F Role-level exposure · p. 27, Table 10 · p. 27

G Ban model

Buff exposure was positively associated with opponents banning buffed champions, while the estimated size varied across specifications and the analyses do not identify causal displacement. The section also reports designated-patch timing and calibration and power results for the outcome test.

G Ban model

  • The ban model uses game-team-champion observations, with pick-share controls and fixed effects; its main interaction specification adds team-specific patch-status and exposure interactions.The analysis excludes unavailable champion choices and retains 258,494 observations.Source: G Ban model · p. 27
  • +11.8 pp was the estimated buffed-champion pick-share interaction in the main exposure-interaction model (95% interval +3.8 to +19.7).The corresponding estimate for other champions was −2.5 pp (−5.2 to +0.1).Source: G Ban model · p. 27
  • Across checks, the buffed-champion interaction stayed positive, ranging from +9.6 to +13.3 pp, while estimates for other champions ranged from −3.3 to −1.3 pp.Excluding 2025, using series game 1, and restricting to first-phase bans also gave positive estimates for buffed champions; the first-phase estimate was smaller, and none identifies causal displacement.Source: G Ban model · p. 28
  • A 1.8-percentage-point lower familiar-pick share was associated with one additional excess ban on buffed champions within team-years (β = −0.018, p = 0.31), not establishing equivalent draft difficulty.Buff exposure had little team-level correlation with familiar-pick share (r = 0.03).Source: G Ban model · p. 28

H Championship patch timing

  • Table 11 compares designated Worlds patch release with event start; the interval was 5–21 calendar days, with a median of 9.5.These dates do not measure each team’s preparation time or include later hotfix timing; exposure analyses instead use team-specific windows.Source: H Championship patch timing · p. 28, Table 11 · p. 28

I Outcome-test calibration and conditional power

  • The studentized test rejected in 5.1% of 2,000 null simulations, while estimated power was 49% at a positive 2.5 pp slope and 83.2% at 3.5 pp.The approximately 3.4 pp 80% power threshold is conditional on the simulation model; the illustrative best-of-five conversion is not a title-probability estimate.Source: I Outcome-test calibration and… · p. 29
Loading 2610.07427v1…