Take Baron2,008 won fights
73.7%raw final win
74.3%73.4%–75.1%
+2557ggold diff · 120s
56.8%structure · 120s
REPLAY EVENT DECISION STUDY · LIVE MATERIALIZATIONWin the fight.
Win the fight.
Then win the map.
28,256 qualified won teamfights. Follow the next 30, 60, and 120 seconds to see what each rank does—and which choices retain the win.
RANK COHORT28,256 won fights
FIGHT WIN STRENGTH · 0–100
How hard did they win?
Death margin anchors the band. Kill share and observed bounty swing refine the score without looking at what the team chose afterward.
28,256 selected fightsShowing decisions across every qualified fight win.
DECISION MATRIX
What happens after the fight ends?
Frequency, final win, and the two-minute state change. Withheld cells failed support or model gates. Nexus sequences are terminal states shown descriptively, never pooled with inhibitor takes or ranked as the best follow-up.
DecisionFrequencyRaw winAdjusted winGold swingStructure
Take Dragon or Elder4,908 won fights
71.9%raw final win
70.2%68.3%–72.4%
+1647ggold diff · 120s
39.2%structure · 120s
Take another neutral objective1,934 won fights
68.3%raw final win
63.8%62.9%–64.5%
+1410ggold diff · 120s
30.6%structure · 120s
Take an inhibitor351 won fights
81.5%raw final win
85.9%association only · no pairwise overlap
+1593ggold diff · 120s
86.6%structure · 120s
Push Nexus turrets or end104 won fights
77.6%raw final win
Withheldterminal state · descriptive
+1094ggold diff · 120s
96.2%structure · 120s
Take a tower11,013 won fights
73.2%raw final win
69.4%67.2%–72.1%
+2340ggold diff · 120s
78.5%structure · 120s
Reset and spend7,938 won fights
52.5%raw final win
57.8%54.9%–60.8%
+917ggold diff · 120s
12.8%structure · 120s
Chase or stagger kills0 won fights
Withheldraw final win
Withheldlow sample
Unavailablegold diff · 120s
Withheldstructure · 120s
Push waves or take map resources0 won fights
Withheldraw final win
Withheldlow sample
Unavailablegold diff · 120s
Withheldstructure · 120s
No measurable conversion0 won fights
Withheldraw final win
Withheldlow sample
Unavailablegold diff · 120s
Withheldstructure · 120s
DECISION CHAIN
The first call is not the whole conversion.
Each window remains attached to the same packet fight so a reset into Baron is not flattened into “Baron,” and a chase that later becomes a tower keeps the chase.
- +30sCommitment
Who leaves, chases, starts an objective, or begins a siege?
- +60sResolution
Gold, XP, deaths, structures, and neutral-objective conversion.
- +120sRetention
Did the team bank the lead, extend it, or give the advantage back?
PAIRWISE WIN IMPACT
Supported local comparisons
Positive values favor the first action. These are local adjusted associations among comparable observed states.
95% interval 2.6 to 7.5 pp
83% propensity overlap95% interval 4.1 to 8.8 pp
80% propensity overlap95% interval -2.3 to 3.8 pp
100% propensity overlap95% interval 9.2 to 16.2 pp
94% propensity overlap95% interval 3.1 to 9.2 pp
94% propensity overlap95% interval 8.1 to 15.4 pp
84% propensity overlapREADING THE RESULT
Fight result first. Decision second. Outcome last.
- Won teamfight
- Fight derived directly from clustered champion-death replay events with at least five participants, at least two deaths, complete surrounding state, and a kill-differential swing of at least two toward one team.
- Fight dominance
- A 0-100 audit score based on the observed death margin, share of fight deaths taken by the losing team, and net kill-bounty swing when at least 80% of deaths carry bounty evidence. Bands remain directly interpretable: close is a two-death margin, decisive is three, and rout is four or more.
- Decision
- First measurable consequential action after fight end within 120 seconds; reset, chase, map pressure, and drift are used only when no milestone resolves first. Inhibitor takes and Nexus pushes are separate treatments. Nexus pushes are terminal-state descriptions and are excluded from adjusted comparisons.
- Win impact
- Adjusted association from five-fold cross-fitted doubly robust models using only pre-fight state. It is withheld when support or overlap fails and is not labeled causal.
- Coverage
- 100.0% rank-resolved · 5,837 matches · minimum 50 observations per action.