These are the exact fragments the model serves — also available as an
ordered study guide.
What probability measures
Probability puts a number on how likely something is, and the number always lives on the same fixed scale: 0 for something that cannot happen and 1 for something that is certain to. Everything else — a coin landing heads, a delayed flight, a sensor misreading a value — sits somewhere between those two ends, and the closer to 1, the more likely.
Making that scale useful takes two pieces of vocabulary that the rest of this concept leans on. The sample space is the complete list of everything that could happen in a given situation — every face a die could show, every card that could be drawn. An event is any subset of that list a person cares about: rolling an even number, drawing a heart, the sensor reads above threshold. Probability is the tool for turning a described event into a single number on the 0 to 1 scale.
None of this requires the outcomes themselves to be numbers, or even similar to each other. A coin flip, a weather forecast, and a sensor reading all fit the same frame: name the sample space, name the event, and ask what fraction of the space the event covers.
Sample spaces and events
The sample space for one die roll holds all six outcomes, 1 through 6; the event rolling an even number is the subset containing 2, 4, and 6.
Every chance process has a sample space: the complete set of outcomes it could produce, with nothing left out and nothing repeated. Roll one fair six-sided die and the sample space is exactly six outcomes, 1 through 6 — no fractional faces, no seventh possibility, no outcome missing. Flip a coin and the sample space shrinks to two outcomes, heads and tails. The sample space is fixed by the process itself, not by which outcome a person happens to be interested in.
An event is any subset of that sample space — like a bag holding every possible outcome, where an event is simply the group of outcomes someone reaches in and pulls out as one bundle. Rolling an even number on a die is the event containing 2, 4, and 6: three of the six outcomes in the sample space, picked out because they share the property being asked about. An event can be as small as a single outcome (rolling exactly a 4) or as large as the entire sample space (rolling a number from 1 to 6, which is certain), and two different events can share outcomes or share none at all.
Naming the sample space and naming the event precisely is the first and most error-prone step in any probability question. Get the sample space wrong — miscount the outcomes, or leave one out — and every probability computed from it is wrong before any arithmetic happens.
The rules probability must obey
On a 0 to 1 probability scale, impossible sits at 0 and certain sits at 1; the disjoint events odd and even on a die each cover half the scale, and their probabilities add to exactly 1.
Probability is not just any number assigned to an event; it has to obey three rules, and every valid probability follows from them. The modern statement of these rules traces to the mathematician Andrey Kolmogorov, who fixed them in 1933: probability can never be negative, the entire sample space has probability exactly 1, and the probability of two events that share no outcomes adds together to give the probability of either happening.
The first rule sets the floor: P(event) >= 0. Nothing can be less likely than impossible, so a probability under 0 is never valid, no matter how the computation got there. The second rule sets the ceiling by fixing the whole sample space at 1: something from the sample space is guaranteed to happen, so the probabilities of every outcome in it, added together, total exactly 1. On a fair die, 1/6 for each of the six outcomes sums to 6 * 1/6 = 1, and that total can never be more or less.
The third rule, additivity, governs events that cannot both happen at once — rolling an odd number and rolling an even number on the same roll, for instance. Odd and even between them cover the whole die, and neither can happen alongside the other, so their probabilities add directly: 3/6 + 3/6 = 1. That single rule is what makes probability arithmetic instead of just labeling: whenever two events share no outcomes, the probability of one or the other is nothing more than addition.
Counting equally likely outcomes
Six equally likely die-roll outcomes in a row; the two favorable outcomes, 5 and 6, are highlighted, giving 2 out of 6 outcomes, or 1/3.
When every outcome in a sample space is exactly as likely as every other, probability collapses into counting: P(event) = favorable outcomes / total outcomes. That single formula, sometimes called classical probability, is the same ratio work already familiar from comparing quantities, just capped so it always lands between 0 and 1.
A fair six-sided die makes every one of its six outcomes equally likely, so counting is all the arithmetic requires. The event rolling a number greater than 4 contains exactly two of the six outcomes, 5 and 6, giving P(greater than 4) = 2/6 = 1/3. A standard 52-card deck works the same way: it holds 13 hearts among its 52 cards, so P(heart) = 13/52 = 1/4, the same favorable-over-total counting, just with a bigger sample space.
The formula only holds when equally likely genuinely applies. A weighted die, a biased coin, or a sample space where some outcomes are simply more probable than others breaks the assumption, and counting outcomes without weighting them would give the wrong answer. Equally likely outcomes is a property of the process being modeled, never something to assume by default.
The complement rule
A circle split into two wedges for one die roll: a 60 degree wedge for rolling a 6, probability 1/6, and the remaining 300 degree wedge for its complement, not rolling a 6, probability 5/6; the two wedges make a full circle.
Every event has a complement: everything in the sample space that is not in the event. Roll one die and the complement of rolling a 6 is rolling anything except a 6 — the other five outcomes, 1 through 5. An event and its complement never overlap, and together they cover the entire sample space, which is exactly the setup the additivity rule needs.
That setup produces a shortcut. Since an event and its complement are disjoint and together fill the whole sample space, their probabilities must add to 1: P(event) + P(not event) = 1, which rearranges to P(not event) = 1 - P(event). Rolling a 6 has probability 1/6, so rolling anything but a 6 has probability 1 - 1/6 = 5/6 without counting a single outcome directly — like a pie cut into two pieces: however big one slice is, the rest of the pie is exactly whatever is left.
The complement rule earns its keep whenever the event itself is awkward to count but its opposite is not. At least one questions are the classic case: counting every way to get at least one success across several tries is tedious, while counting the single way to get zero successes and subtracting from 1 is not. The rule is exact, not an approximation, precisely because it rests on the same two axioms as everything else in this concept.
Combining events, the addition rule
Two overlapping events on one die roll: A is even numbers, B is greater than 4, and they share exactly one outcome, 6, so the addition rule subtracts that shared outcome once.
Combining two events with or takes more care than combining them with and. P(A or B) is not simply P(A) + P(B), because any outcome that belongs to both events would get counted twice — once inside each probability — and a probability can never legitimately count the same outcome more than once.
The fix is the addition rule: P(A or B) = P(A) + P(B) - P(A and B), subtracting the overlap exactly once to undo the double count. Roll one die and let A be rolling an even number (3/6) and B be rolling a number greater than 4 (2/6). The two events share exactly one outcome, 6, so P(A and B) = 1/6, and P(A or B) = 3/6 + 2/6 - 1/6 = 4/6, not the 5/6 a naive sum would give.
When two events share no outcomes at all — mutually exclusive events, like rolling a 2 and rolling a 5 on the same roll — the overlap term is 0 and the formula simplifies back to plain addition, the same additivity already covered for an event and its complement. The subtraction only ever matters when the events can genuinely happen together.
Independent events multiply
A 2 by 2 grid of two independent coin flips, coin 1 as rows and coin 2 as columns; the favorable outcome, heads on both, fills exactly one of the four equal quarters, matching 1/2 times 1/2 equals 1/4.
Two events are independent when knowing the outcome of one tells nothing at all about the other. Flipping two separate coins is the clean case: whatever the first coin shows has zero bearing on what the second coin shows, because nothing physically connects them — like flipping a coin in one room while rolling a die in another: nothing about one result can travel to the other.
For independent events, and becomes multiplication: P(A and B) = P(A) * P(B). Two fair coins both landing heads is P(heads) * P(heads) = 1/2 * 1/2 = 1/4. Two fair dice both showing a 6 is P(6) * P(6) = 1/6 * 1/6 = 1/36, out of the 36 equally likely outcomes that rolling two dice together produces in total. Multiplication, not addition, is the signal that two events are being combined with and rather than or.
Independence is a claim about the situation, not a rule that always applies, and it fails the moment one outcome genuinely changes the odds of another — drawing two cards from the same deck without replacing the first, for instance, where removing one card shifts what remains for the second draw. Sorting out exactly how much one outcome changes another's odds is a question of its own, and a different rule than multiplication answers it.
Where probability shows up
An agent pipeline that calls an external tool cannot know in advance whether that call succeeds, times out, or returns something unexpected — planning for all three means reasoning in probabilities, not certainties, and every rule in this concept is exactly the arithmetic that reasoning runs on. A team tool estimating how long a task will take, or how likely a deadline is to slip, is doing the same thing with a different vocabulary: outcomes, events, and a number between 0 and 1.
Physical systems make the same demands from the other direction. An RC car's distance sensor does not report a single true value so much as a range of plausible ones, and a self-driving build has to combine several such readings, each with its own uncertainty, into one decision about what is actually ahead. None of that combining works without the exact machinery covered here: complements for what if this reading is wrong, unions for either sensor flags a problem, independence for these two failures have nothing to do with each other, and the axioms underneath all of it guaranteeing the arithmetic stays honest.
None of this requires trusting intuition about likelihood, which is notoriously unreliable. It requires naming a sample space, naming an event, and doing the counting and arithmetic this concept lays out — the same handful of rules, however dressed up the surrounding system gets.