Crits, Dragons, and Ganks: Probability on the Rift
Module 2 · Probability

Crits, Dragons, and Ganks: Probability on the Rift

Model critical strikes with independent events and the binomial distribution, see how streak-breaking crit systems create dependent events, reason about dragon spawn order without replacement, and use conditional probability and Bayes' rule to decide when a gank is coming.

Probability Is Long-Run Frequency

A probability is a number between 0 and 1 that describes how often something happens over many repetitions. “25% crit chance” means that over thousands of auto-attacks, about a quarter of them crit. It does not mean every fourth attack crits.

Three rules carry most of this lesson:

RuleFormulaRift example
ComplementP(not A) = 1 − P(A)P(no crit) = 1 − 0.25 = 0.75
AND (multiplication)P(A and B) = P(A) · P(B given A)Win the fight and secure Baron
OR (addition)P(A or B) = P(A) + P(B) − P(A and B)Infernal or Ocean drake spawns first

Independent Events: Crit Chance

Two events are independent when one tells you nothing about the other. If every auto-attack rolls its crit with a fresh coin flip, crits are independent, and the multiplication rule simplifies to P(A and B) = P(A) · P(B).

With 25% crit chance:

  • P(two crits in a row) = 0.25 × 0.25 = 6.25%
  • P(no crit in 6 autos) = 0.75⁶ ≈ 17.8%
  • P(at least one crit in 6 autos) = 1 − 0.75⁶ ≈ 82.2%

That last one uses the complement trick: “at least one” is hard to count directly, but “none” is one easy product.

The Binomial Distribution

If you make n independent attempts, each with success chance p, the number of successes k follows a binomial distribution:

P(k crits in n autos) = C(n, k) · pᵏ · (1 − p)ⁿ⁻ᵏ

C(n, k) (“n choose k”) counts how many different orderings produce exactly k crits. The distribution’s expected value is n · p. For 6 autos at 25%, that’s 1.5 crits per trade.

Expected Damage

With attack damage AD, crit chance p, and crit multiplier m, each auto’s expected damage is:

E[damage] = (1 − p) · AD + p · m · AD = AD · (1 + p · (m − 1))

At 150 AD, 25% crit, and a 1.75× multiplier: 150 × (1 + 0.25 × 0.75) ≈ 178 per auto. Expected value is what you get on average; any single trade can land far from it.


Dependent Events: Streak-Breaking Crits

Pure independence feels unfair to players: sometimes you crit four times in a row, sometimes not once in ten attacks. Many games, League of Legends on PC among them, smooth this with a pseudo-random distribution (PRD). Each attack that doesn’t crit raises the chance on the next one, and a crit resets it:

P(crit on the k-th attack since your last crit) = C · k

The constant C is tuned so the long-run average still equals your listed crit chance. Now the attacks are dependent: what happened last attack changes the odds of this one. Over many fights the average is the same, but the spread (variance) is smaller, so there are fewer zero-crit trades and fewer four-crit spikes.

Whether a specific game or mode uses PRD for a given mechanic is a detail to verify, not assume. The lab lets you compare both models.

🎮 Rift Lab: Critical Strike Simulator

Bars show the exact binomial probability of landing k crits in your trade. Hit Simulate to roll 5,000 trades and compare independent rolls against a pseudo-random (streak-breaking) crit system with the same average chance.

P(≥ 1 crit)—
P(0 crits)—
P(2 crits in a row)—
Expected crits—
Expected dmg / auto—
Expected trade dmg—
Not simulated yet.
Number of crits in one trade
Exact binomialSimulated: independentSimulated: pseudo-random

Sampling Without Replacement: Dragon Order

Suppose (as a simplified rule) the first elemental drake is equally likely to be any of four types (Infernal, Mountain, Ocean, Cloud), and the next drake can’t repeat the previous type. The second drake is now drawn from three options, not four. That’s sampling without replacement, and it makes the events dependent.

P(Infernal is one of the first two drakes):

  • Infernal first: 1/4
  • Not Infernal first (3/4), then Infernal from the remaining three (1/3): 3/4 × 1/3 = 1/4
  • Total: 1/4 + 1/4 = 1/2

If drakes were independent with repeats allowed, the answer would be 1 − (3/4)² = 7/16 ≈ 0.44. The no-repeat rule raises the chance of seeing any specific type early.

P(the first two drakes are both Infernal) is 1/16 under independence but 0 under the no-repeat rule. When events are dependent, multiplying the individual probabilities gives the wrong answer.

Real spawn rules (which drake types exist, when the map changes, when Elder Dragon appears) vary by game and patch. The point is to recognize which model a rule implies.


The Law of Total Probability: Baron Calls

You’re deciding whether to start Baron. Break the outcome into cases:

  • P(your team wins the fight at Baron) = 0.6, and if you win, you secure it 95% of the time.
  • If you lose the fight (0.4), there’s still a 10% chance someone steals it.
P(secure) = P(win) · P(secure | win) + P(lose) · P(secure | lose)
          = 0.6 · 0.95 + 0.4 · 0.10
          = 0.57 + 0.04 = 0.61

This is the law of total probability: split into cases that cover every possibility, then weight each case by how likely it is.


Conditional Probability and Bayes’ Rule: Ganks

P(A | B) (read “A given B”) is the probability of A when you already know B happened:

P(A | B) = P(A and B) / P(B)

Top laners constantly face this question: the enemy jungler just disappeared from the minimap. How likely is a gank?

Let G = “a gank is coming” and E = “the jungler is missing from vision.” Your experience gives you:

  • Base rate P(G) = 0.15: in a typical 30-second window, a gank comes 15% of the time.
  • P(E | G) = 0.9: before a gank, the jungler is almost always missing.
  • P(E | not G) = 0.3: the jungler is also often missing when they’re just farming.

Bayes’ rule flips the conditional:

P(G | E) = P(E | G) · P(G) / P(E)

P(E) = 0.9 · 0.15 + 0.3 · 0.85 = 0.135 + 0.255 = 0.39

P(G | E) = 0.135 / 0.39 ≈ 0.35

Seeing the jungler go missing raises the gank chance from 15% to about 35%, a 2.3× jump. And if the jungler is visible, P(G | not E) = 0.015 / 0.61 ≈ 2.5%. Vision is valuable because it changes your probabilities in both directions.

🎮 Rift Lab: Gank Probability with Bayes' Rule

You're in top lane. Should you back off when the enemy jungler goes missing? Each square below is one of 100 similar moments from your match history.

P(gank | missing)—
P(gank | visible)—
P(missing)—
Risk multiplier—
Gank, jungler was missing Gank, jungler was visible No gank, jungler was missing (false alarm) No gank, jungler visible

⚠️ Common Pitfall: Ignoring the Base Rate

Common mistake: “The jungler is missing before 90% of ganks, so a missing jungler means a 90% gank chance.” That confuses P(E | G) with P(G | E). If ganks are rare (low base rate) and junglers go missing for many other reasons, most “missing” moments are false alarms. The same error shows up in medical tests, fraud alerts, and cheat detection: a 99%-accurate detector can still be wrong most of the time it fires when the thing it detects is rare.

🧠 Knowledge Check

1. With 20% crit chance and independent rolls, what is P(at least one crit in 5 autos)?

2. A streak-breaking (pseudo-random) crit system with the same average crit chance changes what?

3. In Bayes' rule, what does lowering P(E | not G), the false alarm rate, do to P(G | E)?