Expected goals explained: what xG actually measures

A player hits one from twenty-five yards, it goes in off the underside of the bar, and the graphic says the shot was worth 0.04. Somewhere in a studio, within about nine seconds, a former professional explains that this proves expected goals is rubbish, because the ball is in the net and the number said it would not be.

The number said nothing of the kind. Expected goals, explained precisely, is a statement about a population of shots rather than about this one: 0.04 means that attempts of that type go in roughly four times in a hundred, and that this was one of the four. A model that priced the shot at 0.04 and never saw one go in would be broken. The strike was excellent and the number was right, and there is no tension between those two statements.

Almost every argument about xG traces back to one misreading, and the fix is a single sentence.

A shot's xG is not a prediction about that shot. It is the base rate for shots like it.

Everything worth knowing follows from that, including all of the things the measure genuinely cannot do.

What expected goals actually measures, and what it does not

Take every shot in a large historical archive. Sort them by the things you can record about the situation: how far from goal, at what angle, with which body part, arriving from a cross or a through ball or a rebound, with the goalkeeper here and two defenders there. Find the group of past shots that most closely resembles the one in front of you. Count how many of them were scored. That share, expressed between zero and one, is the shot's expected goals value.

The word "expected" is doing statistical work, not everyday work. It means the average outcome over the reference class, in the sense that a fair die has an expected value of 3.5 despite never once landing on 3.5. It does not mean anticipated, deserved, or likely. A shot worth 0.30 is not a shot that ought to have gone in. It is a shot from a family in which roughly three in ten went in.

That distinction matters because of what it removes from the number. A shot's xG contains no information about the striker's ability, none about how cleanly the ball was struck, and none about where in the goal it finished. The model assigns its value at the instant of contact, from the situation alone. Whether the shot then screams into the top corner or dribbles into the goalkeeper's shins is outside the calculation entirely.

Three consequences worth stating plainly.

The same shot always gets the same number. If the world's best striker and a centre-half caught upfield take identical shots from identical positions with identical defensive pressure, a standard model gives them identical values. This upsets people, and it is deliberate. A model that adjusted for shooter identity would be measuring the shooter rather than the chance, and the entire point of the exercise is to describe the chance so that the shooter can be judged against it afterwards.

A single shot's value can never be tested. There is no experiment in which a 0.12 shot is taken a thousand times under identical conditions. The number is only checkable in aggregate: gather ten thousand shots the model priced near 0.12 and see whether around twelve hundred were scored. That is how a model is calibrated, and it is the only sense in which xG can be right or wrong at all.

Summing probabilities gives an expected count, not a likely score. This is where broadcast graphics do the most damage. A team that finishes a match on 1.0 xG did not create one goal's worth of chances in the sense of being on course for one goal.

Work it through with an invented example. A side takes five shots, each worth 0.20, so its xG total is exactly 1.0. Treat the shots as independent, which is a simplification the last section of this article returns to. The chance of scoring none of them is 0.8 multiplied by itself five times, which is about 33 per cent. The chance of exactly one is about 41 per cent. Two or more comes to about 26 per cent.

Scores no goals32.8%

Scores exactly one41%

Scores exactly two20.5%

Scores three or more5.8%

An invented team taking five shots worth 0.20 each, treated as independent. Percentages are the binomial arithmetic on those invented values, rounded. No real match, team or season is described here.

Show the numbers

Worked example: what a total of 1.0 xG actually delivers

Item	Value
Scores no goals	32.8%
Scores exactly one	41%
Scores exactly two	20.5%
Scores three or more	5.8%

Exactly one goal is the single most likely outcome and it still happens well under half the time. Scoring nothing at all from a full 1.0 xG happens about a third of the time, which is not bad luck, an outrage, or evidence of a finishing crisis. It is the arithmetic working normally.

Expected goals explained: how one shot becomes a number

The features a model reads are unglamorous, and the ordering of their importance is stable across almost every implementation.

Distance to goal. The single strongest input, and not in a straight line. The drop-off in scoring rate as you retreat from the six-yard box is steep at first and then flattens out, because past a certain range almost everything is hard and the difference between hard and slightly harder is small.

Angle. Covered properly in the next section, because it is the feature that most people quietly conflate with distance and it is not the same thing at all.

Body part. Foot, head, or something else. A header from a given spot is worth materially less than a foot shot from the same spot, because the shooter controls neither placement nor power to anything like the same degree, and because headers overwhelmingly arrive from crosses that the defence has had time to organise against.

How the ball arrived. This is the feature that separates a serious model from a crude one. A shot after a through ball tends to face a scrambling defence and an advancing goalkeeper. A shot after a cross tends to face a set defence, a shooter moving across the ball, and a first-time contact with poor control. A rebound is the most valuable arrival of all, because by definition the goalkeeper is already committed and on the floor. Same distance, same angle, three different problems.

Pattern of play. Open play, direct free kick, corner, penalty. Penalties are the extreme case: every one is taken from the same mark, eleven metres from the goal line, with no defender permitted inside the area, which makes them the only shot in football that repeats identically. Every model gives every penalty the same number for exactly that reason.

Defensive pressure and goalkeeper position, where tracking data allows. The best modern datasets capture a freeze frame at the moment of the shot: how many defenders sit inside the triangle between the ball and the two posts, how far off his line the goalkeeper is, and whether the shooter has anyone within a stride of him. A model with that information can distinguish a shot from the penalty spot with six bodies in the way from an identically located shot with a clear sight of goal. A model without it treats them as the same chance, which is one of the largest single reasons two providers disagree.

Everything else. Big-chance flags, whether the ball was bouncing or rolling, whether the shooter took a touch first, whether the shot came from an opponent's error, the game state, the assisting player's foot. Implementations differ in which of these they use and how much weight each carries.

1. 1A shot is loggedAn analyst or an automated tracker records that a player has deliberately attempted to score, with a timestamp and a pitch coordinate for the point of contact.
2. 2The geometry is derivedFrom that coordinate come two numbers: the distance to the centre of the goal line, and the angle subtended by the two posts from where the ball was struck.
3. 3The manner of the shot is codedBody part, whether the shooter took a touch first, whether the ball was on the ground or in the air.
4. 4The build-up is attachedHow the ball got there. Through ball, cross, cut-back, rebound, corner, direct free kick, penalty, a defensive error, a dribble past the last man.
5. 5The picture around the ball is capturedWhere tracking or freeze-frame data exists, the model reads the goalkeeper's position and every defender inside the shooting triangle. Where it does not, this step is skipped and the model is poorer for it.
6. 6The reference class is looked upThe trained model returns the share of historical shots matching that description that were scored. This is a lookup against the past, not a judgement about the present.
7. 7The value is written against the shotThat share is the shot's xG. It belongs to the situation, and it would be identical had a different player taken it.
8. 8The shot joins the running totalsTeam xG, player xG, non-penalty xG, xG conceded. Only in aggregate does the number begin to behave like a measurement of anything.

The route from an event on the pitch to a value in a table. No figures are asserted at any step; the point is the order of operations.

Notice what the flow does not contain. Nowhere does the model consult the outcome, the shooter's reputation, the league table, or the crowd. It also has no view on whether taking the shot was the right decision, which is a limitation the tactics section returns to.

Why the goal is a triangle, not a distance

Distance gets all the attention. Angle does most of the interesting work, and it explains several results that otherwise look perverse.

The Laws fix the goal at 7.32 metres wide between the posts and 2.44 metres high. From any point on the pitch, the goal presents a certain visible width, and the useful way to think about that is the angle between lines drawn from the ball to each post. Stand on the penalty spot and that angle is generous. Walk twenty metres sideways towards the touchline while staying the same distance from the goal centre and the posts close towards each other until the goal is a slot.

This is why two shots from the same distance can be worth wildly different amounts. It is why a ball squared back to the penalty spot is worth more than a ball fizzed across the six-yard box: the first arrives where the goal is wide open, the second where the shooter is shooting across an angle that is closing fast. It is why coaches drill the cut-back until players resent them for it.

Angle also explains the asymmetry of goalkeeping. A goalkeeper covering a shot from a tight angle only has to protect a narrow strip and can position himself to reduce it further. A goalkeeper facing a shot from a central position has to cover the whole frame. The goalkeeper's own position is a feature in its own right in the better models, and it is the input that produces the biggest surprises: a chance that looks routine from the halfway line becomes very valuable indeed if the goalkeeper is caught two metres off his line and moving the wrong way.

The general lesson is that xG is a geometry problem before it is a football problem. Distance, angle, and how much of the frame is obstructed account for the bulk of the variation. The rest is refinement.

Why two providers give different xG for the same shot

Open two websites after a match and the totals will not match. This is treated as a scandal roughly once a season. It is nothing of the sort, and understanding why is a genuinely useful skill.

The training samples differ. One model may be built on several seasons across a dozen leagues; another on one competition. Scoring rates from comparable positions differ between a professional men's top division, a women's top division and a fourth tier, because defensive speed, goalkeeping standard and shooting power all differ. A model trained on one and applied to another will be systematically off.

The feature sets differ. A model with freeze-frame defender positions can see that a shot from eight metres had four bodies between the ball and the goal. A model without it cannot, and will happily price that block-fest as a golden chance. The gap between the two is largest exactly where it matters most, in the crowded area in front of goal.

The event definitions differ. Providers do not agree on what a shot is. Blocked shots, shots deflected wide, attempted crosses that drift towards the net, and clearances that strike an attacker all sit in grey areas. Whether a shot is attributed to the player who struck it or to the deflection that redirected it is a coding rule, not a fact about football. Two identical passages of play can therefore produce a different number of shots before any modelling starts.

The model class differs. A logistic regression on a handful of features produces smooth, conservative values. A gradient-boosted tree ensemble on forty features produces sharper, more confident ones and can pick up interactions the regression flattens. Neither is automatically better; they are different bets about how much structure the data supports.

The calibration target differs. Some implementations are tuned so that total xG across a competition lands close to the total goals actually scored in it. Others are not, and drift slightly high or low as a result.

None of this makes xG unreliable. It makes it an estimate, which is what it always claimed to be. There is a practical use for the disagreement, too. When two competent models give nearly the same value, the chance was an ordinary one that sits in the dense part of the historical data. When they diverge sharply, the chance was unusual, the reference class was thin, and both numbers deserve less confidence than either presents. Disagreement is a crude uncertainty estimate that the models themselves rarely publish.

The one rule that follows is unbreakable. Never mix providers inside a single comparison. A striker's xG from one source set against a team's xG from another is not a comparison of anything.

What xG is genuinely good for

Three things, and they are worth the entire apparatus.

Separating process from outcome across a season. Football is a low-scoring game in which a single deflection redistributes three points. Results are therefore a noisy record of performance, and a league table after a dozen matches is partly a record of which teams have been fortunate. Chance creation and chance concession are far steadier from month to month than conversion is. A side whose xG difference is strongly positive while its goal difference is negative is not being unlucky in a mystical sense; it is doing the repeatable part of the job well and the volatile part badly, and the volatile part tends to revert.

This is the same logical move that golf made when it replaced its box score with a baseline measure: price the situation first, then judge the execution against it. Cricket does it too, and for the same reason, since a batting average conceals which deliveries a player actually faced in exactly the way a goal tally conceals which chances a striker was handed.

Spotting unsustainable runs in both directions. A striker scoring far more than his chances warrant is the story everyone notices. The mirror case is more valuable and almost nobody covers it: a player getting into excellent positions repeatedly and converting poorly. The first player is expensive and about to decline. The second is cheap and about to improve, provided the chances keep arriving, and it is the chances arriving rather than the conversion that carries the signal.

Evaluating chance creation independently of who finished it. Before xG, a creative midfielder's output was measured in assists, which is to say it was measured by whether somebody else scored. Two identical passes into the same square metre of grass produced either a career-defining number or nothing at all depending on a striker the creator did not choose. Expected assists severs that dependency, and it is the single largest improvement xG made to how anyone is evaluated.

The same reasoning has crossed into other sports. Ice hockey runs the identical construction, and while the shot model that hockey uses differs in its features, the definition and the misreadings around it are word for word the football ones.

What xG is bad for, starting with single matches

The measure has clear failure modes. Anyone quoting it should be able to recite all five.

Single matches. A match contains something in the region of twenty to thirty shots between both sides, most of them worth very little. That is a tiny sample, and the totals bounce around accordingly. Worse, the same total can be reached in ways that mean completely different things.

Two invented sides both finish on 1.5 xG. One took three shots worth 0.5 each. The other took fifteen worth 0.1 each. Treating shots as independent again, the first side scores at least once about 87.5 per cent of the time, the second about 79.4 per cent.

• Three shots at 0.50
• Fifteen shots at 0.10

Scores no goals12.5%20.6%

Scores exactly one37.5%34.3%

Scores two or more50%45.1%

Two invented sides, both totalling 1.5 xG. Team A takes three shots at 0.50; Team B takes fifteen at 0.10. Percentages are binomial arithmetic on those invented values, treating shots as independent.

Show the numbers

Worked example: two routes to 1.5 xG in one match

Item	Three shots at 0.50	Fifteen shots at 0.10
Scores no goals	12.5%	20.6%
Scores exactly one	37.5%	34.3%
Scores two or more	50%	45.1%

Same headline number, different match. The side hoarding low-value attempts also tends to be the side chasing the game, which brings us to the next problem.

Game state contaminates everything. A team two goals ahead with twenty minutes left stops trying to create and starts trying to prevent. Its xG accumulation collapses, not because it has become worse but because the task has changed. Meanwhile the losing side piles bodies forward and takes speculative attempts it would never take at 0-0. Raw match xG therefore systematically flatters the team that is behind, and a full-time total of 2.1 against 0.6 may describe forty minutes of desperation rather than forty minutes of superiority. Serious analysis splits xG by game state before drawing conclusions. Broadcast graphics never do.

Penalties distort totals out of all proportion. A penalty is worth several times an average open-play attempt, and awarding one is heavily influenced by refereeing and by video review. A team with three penalties in five matches has an xG total that says almost nothing about how it is playing. This is not a hypothetical distortion; it is the most common single reason a total misleads, and it is why non-penalty xG exists. The point is sharpened by the fact that a chunk of modern penalty awards come from incidents a referee did not see live and was sent to review, which means the input to a team's attacking numbers is partly a function of officiating protocol.

Finishing skill needs an enormous sample. This is the failure people find hardest to accept, so it is worth doing the arithmetic.

Invent a striker who takes 100 shots in a season, each worth 0.10. His expected total is 10 goals. The standard deviation of that count, from chance alone with no variation in his ability whatsoever, is the square root of 100 times 0.1 times 0.9, which is 3.

• 100Shots in the invented season
• 10Expected goals at 0.10 a shot
• 3One standard deviation from chance alone, in goals
• 400Shots needed to halve that noise as a share of the total

An invented striker taking 100 shots worth 0.10 each. Every figure is arithmetic on those invented inputs and describes no real player.

Three goals is a third of his expected haul. A player finishing on 13 and a player finishing on 7 are, on this evidence, indistinguishable, and both are perfectly ordinary outcomes for the same average finisher. Since the noise shrinks with the square root of the sample, cutting the relative uncertainty in half takes four times the shots, which is several seasons of a first-team career. Any confident statement about a finisher built on one season's worth of attempts is a statement about noise wearing a lab coat.

It prices outcomes, not decisions. The model has no opinion on whether the shot should have been taken. A player who shoots from a hopeless position when a teammate was free in the middle collects a small positive number, and his team's xG total goes up. Shot selection is a skill, and it shows up in xG per shot rather than in xG, which is one of several reasons those two numbers should always be read together.

The measures built on top, and what each one adds

The base measure spawned a family. Each member exists to answer a question the original could not.

Non-penalty xG (npxG) is the total with penalties stripped out. Use it for essentially every comparison between players and between teams. Penalty-taking is a specialism, penalty-winning is partly luck and partly refereeing, and neither belongs in an assessment of open-play attacking.

xG per shot divides total xG by shots taken. It measures chance quality rather than chance volume, and it is how you tell the difference between the two invented sides in the chart above. A high figure means a team or player only shoots from good positions. A low one means volume from range, which is a defensible strategy against a deep block and a poor one everywhere else.

Post-shot xG, sometimes called xG on target, is the one most often explained backwards. It is computed only for shots that hit the target, and it adds one feature the ordinary model refuses to look at: where in the goal frame the ball was heading. A shot arriving in the bottom corner is much harder to save than the identical shot arriving at the goalkeeper's midriff, and post-shot xG prices that difference.

Because it prices placement, post-shot xG belongs primarily to the goalkeeper. Sum a goalkeeper's post-shot xG faced and subtract the goals he actually conceded, and you have a shot-stopping measure that adjusts for the difficulty of what came at him, which no save percentage ever did. A goalkeeper who conceded fewer goals than the post-shot values suggest has made saves.

Attributing it to the striker is where people go wrong. Post-shot xG only exists for shots on target, so a player who blazes half his attempts into the stands is rewarded by having those attempts excluded from the measure. It is a description of the shots that arrived, not of a player's finishing, and the two are different quantities.

Expected assists (xA) takes the xG of the shot a pass produced and credits it to the passer. A pass that sets up a chance worth 0.4 earns its passer 0.4 in xA whether or not the shot went in. Two clarifications people miss. It is only awarded when a shot actually followed, so a perfect pass that a striker fails to shoot from earns nothing. And it prices the outcome of the pass rather than the intent, so a scuffed clearance that falls kindly earns the same as a defence-splitting through ball to the same spot.

Expected threat (xT) was built to fix that last gap. Instead of valuing shots, it values possession of the ball in every part of the pitch, which means it can put a number on a pass or a carry that never leads to a shot at all.

1. 1Divide the pitch into a gridA few hundred zones, fine enough that a pass usually crosses at least one boundary.
2. 2Measure what happens in each zoneFrom historical data, how often possession in that zone ends in a shot on the very next action, and how often that shot is scored.
3. 3Measure where the ball goes insteadFor the times possession is moved rather than shot, the distribution of destination zones for the pass or carry.
4. 4Solve the whole grid at onceA zone's value is the chance of shooting from it multiplied by the chance of scoring, plus the chance of moving the ball multiplied by the value of wherever it goes. Every zone's value therefore depends on every other zone's, so the system is solved together rather than filled in one square at a time.
5. 5Read off the threat of each zoneThe solved number is the probability of scoring within the next few actions, given the ball is here. It counts moves that lead to moves that lead to shots.
6. 6Value an action as a differenceA pass or a carry is worth the threat of the destination minus the threat of the origin. A ball moved into a more dangerous square is a positive action.
7. 7Credit the player who moved itThe midfielder who slid the ball between the lines gets paid for it even though the move broke down two passes later.

The construction of a possession-value grid. No values are asserted; the point is that the grid is solved rather than counted.

Expected threat, and the possession-value models that generalise it, is how a defensive midfielder who never shoots and rarely assists becomes measurable at all. It is also the natural companion to pressing measures: the same grid that says where the ball is dangerous says where winning it back is valuable, which is the argument underneath the metrics that try to quantify how aggressively a side presses.

Regression to the mean, and the players who refuse to regress

Here is the standard analytical position. A player who has scored well above his xG over a run of matches is expected to score closer to his xG in future, because the run was mostly chance and chance does not persist. This is regression to the mean, it is correct on average, and it has an excellent forecasting record.

It is also stated far too smugly, and the honest counter-case deserves proper space.

Some players beat their xG persistently. Not for ten matches, for hundreds of shots across many seasons, at a rate that the binomial arithmetic above struggles to explain. When that happens, the intellectually honest conclusion is not that the player has been lucky for six years. It is that the model is missing something about how that player shoots.

Several things could be missing, and all of them are real.

Placement, which the base model refuses to look at. A player who consistently finds the corners is generating outcomes the ordinary model cannot see, because the ordinary model stops at the moment of contact. Post-shot xG catches part of this, and only for shots on target.

Shot timing within the situation. Two players in nominally the same position are not in the same situation if one shoots half a second earlier, before the defender closes and before the goalkeeper sets. Event data records where the shot was taken from, and not always what the defence was about to do.

Technique that beats the reference class. Power on a first-time strike, disguise in the shaping of the foot, the ability to hit a ball cleanly off a bad bounce. The reference class is built from everybody's shots, and everybody includes a great many players who cannot do those things.

Systematically mispriced chance types. This applies to whole teams. A side that generates its shots through a specific repeatable pattern, a particular cut-back after a particular press trigger, may be creating chances whose real conversion rate is higher than the model's reference class implies, because the reference class averages over hundreds of superficially similar shots created much worse. Set-piece routines designed around a specific delivery and a specific run are the clearest example, since a coached routine is a different animal from the general population of corners that trained the model.

So the correct response to persistent overperformance is to improve the model, not to declare the evidence inadmissible. Model builders know this, which is why shot models have grown steadily more detailed and why freeze-frame features exist at all.

Two things keep regression as the default expectation anyway.

The first is base rates. Genuine, durable overperformers are rare, and short runs of overperformance are extremely common, so a random overperforming run is far more likely to be the second than the first. Betting on regression will be right most of the time by construction.

The second is the identification problem, which is brutal. Distinguishing a real finishing edge from noise takes so many shots that by the time the evidence is conclusive, the player is usually past his peak. The information arrives too late to act on, which is precisely why recruitment departments weight chance quality and chance volume heavily and finishing lightly. It is not that they think finishing does not exist. It is that they cannot measure it in time to buy it.

There is a survivorship trap under all of this too. The strikers whose names come up in this argument are the ones who overperformed, because those are the ones who stayed in the team long enough to accumulate a career. The players who underperformed an identical set of chances by an identical margin, purely by chance, were dropped at twenty-three and are not available for the debate.

What xG did to recruitment, and to the way matches are played

The practical effects have been larger than the analytical ones, and mostly invisible from the sofa.

In recruitment, the shift was from goals to the components of goals. A striker's tally is a joint product of the chances he received, the system that produced them, and his conversion. Two of those three do not travel with him when he changes clubs. Pricing a player on chances created and chances received, rather than on the goals column, lets a club distinguish the forward who was carried by a system from the forward who created his own value. That is how a market inefficiency opened up, and it is why clubs across the professional football pyramid now employ people whose job is to argue with the scouting department using shot maps.

Two specific patterns emerged. Clubs learned to sell strikers coming off large overperforming seasons, because the buyer pays for the goals and receives the chances. And they learned to buy creators whose xA has run well ahead of their assists, because the assist column is a tax on having played with poor finishers.

In-game, the effects are visible if you know where to look. Speculative shooting from distance declined in the sides that took the numbers seriously, because the arithmetic on a thirty-yard effort against a set defence is unforgiving. Cut-backs rose, because the geometry section explains why they are worth so much more than the same ball delivered flat across the six-yard box. Corner routines were rebuilt around producing a specific contact in a specific zone rather than any contact anywhere. Attacking players learned to take an extra touch to improve the angle rather than shooting first time from a bad one.

The counter-arguments are real and worth stating.

Coaching to a metric produces predictable teams. A side that never shoots from distance can be defended by dropping deep and conceding the areas that carry no threat, which is exactly what several teams started doing in response. The long shot has some value as a threat that keeps defenders honest even when its direct expected return is poor, and a model that only prices the shot itself cannot see that second-order effect.

There is also a positional blind spot. A forward who drops between the lines to create rather than occupy the last defender will post modest shot numbers by design, which is why the tactical role that trades shots for space looks like an underperforming striker to anyone reading only an xG column. The measure is not wrong. It is answering a question about shots when the question being asked is about structure.

Reading a broadcast xG graphic without being had

The graphic at full time gives you two numbers and no context. Six questions restore the context, and they take about fifteen seconds.

Whose model is it? Different providers, different numbers. If the broadcaster does not say, the figure cannot be compared with any figure you saw elsewhere this week.

Does it include penalties? One penalty can be most of the gap between the two totals. If the graphic does not distinguish, and it almost never does, mentally deduct penalties from both sides before you conclude anything.

Is it xG or post-shot xG? A rising number of graphics show post-shot values without labelling them, because they look more dramatic. Post-shot totals exclude every shot that missed the target, so they describe accuracy as much as chance creation and they are not comparable with ordinary xG.

How many shots produced it? A total of 1.8 from two shots and a total of 1.8 from eighteen are completely different matches, and only the shot count tells you which one you watched. If a shot map is available, look at the shape rather than the sum.

Was one shot most of the total? A single big chance can carry a whole match's figure. If it did, the total is really a statement about one moment rather than about ninety minutes of play, and the losing manager complaining about xG may have an entirely fair point.

What was the game state? A total accumulated while chasing a two-goal deficit measures desperation. Split the match at the first goal in your head and ask whether the numbers still say what the graphic implies.

Then apply the discipline the whole measure asks for. One match of xG is a small sample of an already noisy process. It is evidence, it is more informative than possession percentage or shots on target, and it is nowhere near enough on its own to overturn what happened on the pitch. The team that scored more goals won. Expected goals never disputed that and was never built to.

What it does is audit the chances. It tells you what a side generated, what it allowed, and how far the result travelled from the process that produced it. Watch that gap across thirty matches rather than one and it stops being an argument and becomes a measurement, which is all the number was ever asking to be.