The Forgotten Label Layer: Semi-Automated Offside and the Unaudited Part of Every Phase

On 1 December 2026, minute 51, Khalifa International Stadium. Kaoru Mitoma sprinted to the byline, kept the ball in play at the last possible instant, and cut it back. Ao Tanaka tapped into an empty net. The assistant referee had already raised his flag. The Japanese end held its breath, then the big screen lit up with a three-dimensional animation: the ball was still in play by exactly 1.88 mm. The goal stood. Spain advanced on goal difference. Germany went home.

A unit smaller than the thickness of a sheet of paper, and it split the fates of two football nations. In the days that followed, people argued about the law, about feeling, about machines having taken something from the game. Very few argued about what actually produced that moment: the way a phase of play gets labelled.

In June 2026, in Bloemfontein, Frank Lampard struck the crossbar and the ball dropped nearly a metre over the line. The referee did not give it. Four days later the whole world argued about an error the human eye could not process at that speed. It took until July 2026 for IFAB to formally approve goal-line technology; the 2026 World Cup in Brazil was the first major tournament to run it across every match.

Since then, each cycle has added another layer of machinery. 2026 brought VAR, with video banks and screen referees. 2026 brought semi-automated offside: twelve cameras mounted under the stadium roof, tracking twenty-nine points on each player's body, sampled fifty times per second, plus a sensor inside the ball transmitting five hundred times per second. Euro 2026 ran the same system. And the 2026 World Cup — forty-eight teams, one hundred and four matches — will be the largest classification workload ever loaded into a single tournament.

Every one of those steps was sold with the same promise: fewer errors. Look closely at the architecture of a decision, and that promise covers only one third of the job.

An offside decision has three layers, and the technology has only touched the first.

The first layer is geometry: who is where, at what instant, which body part counts. Machines do this better than human eyes, and far better. The 1.88 mm belongs to this layer.

The second layer is semantics: what kind of phase this is. The offside law does not speak in millimetres. It speaks of "involved in active play", "interfering with an opponent", "deliberate play". Those are labels, and labels are assigned by people, even after the machine has drawn the line.

The third layer is context: what that label is then used for in a larger story — group tables, goal difference, qualification, contracts, transfer fees, continental places.

Based on my experience tracking matches played under semi-automated systems, most of the argument in the stands and online sits in the second layer, while the evidence the two sides throw at each other always sits in the first. A three-dimensional frame accurate to the millimetre is used to rebut an opinion that is really about classification. Two people speak two languages and both believe they are discussing the same thing.

In the second layer, what decides the outcome is not a camera but a dictionary. Who wrote that dictionary, when, and whether it was revised mid-season — those are the least-asked questions in the entire football analysis industry right now.

Even the line has its own dictionary. Current guidance measures from the bottom of the armpit, and that convention has shifted from season to season. One goal can be ruled out by a shirt sleeve; another can stand because the same body part fell on the other side of a line thinner than a hair. Same player, same posture, two opposite conclusions — separated by which month the guidance was updated.

Then there is the label "clear and obvious error". It is a phrase with no quantity attached. No threshold, no unit, no conversion table. A three-person VAR team can apply that label to the same phase in three different ways, and there is no mechanism to say who is right, because the label itself does not permit verification.

In statistics the problem is wider. Same phase, same name, two providers, two definitions. "Successful tackle" at one provider includes the case where a player deflects the ball and a teammate recovers it; the other counts only the case where the player wins the ball himself. "Progressive pass" depends on whether you measure along the pitch's vertical axis or the actual direction of attack. "Big chance" depends on location, number of defenders, and whether a goalkeeper is present.

I learned this rather late. In 2026, when the A-League reached round 12, I published a 3,200-word analysis of Melbourne City's pressing metrics using GPS positional data. The conclusion: Warren Joyce's side pressed in the wrong direction, forcing Luke Brattan to cover 11.2 km per match while producing only 1.3 successful tackles. The piece was mocked for being too dry. Three weeks later Joyce changed the pressing structure and Melbourne City won four in a row.

The part I never wrote about was what happened before that. I used provider A's label, then cross-checked it against provider B's identically named label. The 1.3 I published was correct under A's dictionary. It would have been 1.7 under B's. The conclusion of the piece did not change, but the error band did, and I never said so.

Most errors in football analysis are not in the arithmetic. They are in the object being mislabelled before any arithmetic begins.

A turnover in the second line can be tagged "systemic defensive failure" or "individual error", depending on who is tagging. Two labels lead to two sets of metrics, two coaching conclusions, and ultimately two different transfer decisions. The dictionary does not merely describe the match; it shapes what gets bought and sold after it.

Once, on a consulting project in Sydney, a data file was mislabelled at the intake stage. The label was applied by someone who did not read the content, only the headline. Three weeks later my model produced results that looked entirely plausible and were completely meaningless, because it was comparing two things that were not the same kind of thing. No algorithm caught it. Only a person sitting down to re-read the original headline caught it.

"Before you trust a number, ask where it was born."

In 2026 I wrote a piece predicting Croatia would reach the semi-finals, based on Luka Modrić's chance creation in the group stage. A group of amateur coaches on Reddit called it the work of a bookworm who did not understand football. Croatia reached the final. "In 2026 they laughed at my xG. This year they ask me what xG is."

But once xG became common language, the question shifted from "what is xG" to "whose xG". Two models applied to the same shot typically differ by fractions of a percent; from narrow angles the spread is much wider. And almost no graphic circulating online states which model it used.

"Transfer value is the story, but data is the signature."

In August 2026, PSG paid 222 million euros for Neymar, the highest fee in history at that point. That number has been explained through commerce, image rights, institutional pressure. But behind every major transfer sits a data table, and that table is built from labels defined by people. When a club buys a midfielder because of a high progressive-pass rate, it is buying one provider's dictionary, not a universal property.

In recent years I have noticed a repeating pattern. A graphic appears after a match and spreads faster than any written analysis. It usually has one column, one number, one red arrow. A week later the data provider issues a correction note: the definition changed from a certain round, or the label was reassigned after video review.

The correction note is almost never shared. The wrong number travels; the correction stays put.

This structure is familiar from other fields: a claim spreads, an authority denies it, and the correction only reaches the most measurable part of the original claim. In football, the most measurable part is always geometry. The least measurable part is always semantics. Which is why corrections always look reasonable and are never sufficient.

The 2026 World Cup sets this problem at an unprecedented scale: forty-eight teams, one hundred and four matches, sixteen cities, three countries. The number of phases needing classification grows arithmetically; the number of people writing the dictionary barely grows at all. When volume exceeds manual audit capacity, providers are forced to automate labelling. Automated models learn from old data, and old data carries old bias.

The counterintuitive angle here is not "technology is ruining football". That line is old and it is also wrong. Technology has fixed errors the human eye could never fix, and anyone who has watched a wrongly disallowed goal on replay knows it.

The counterintuitive angle is this: technology does not erase error, it relocates error to a higher tier, into a place nobody audits. When the decision lived in the referee's eye, the error lived in the referee's eye, and everyone could see it. When the decision moves into the camera, the error moves into the definition, and almost nobody sees it, because definitions do not appear on the big screen.

There is a second consequence: confidence in the geometric layer bleeds into the semantic layer. Because 1.88 mm is right, people assume the whole decision is right. It is a logically invalid inference and a psychologically irresistible one, especially in a match where spectators have seventy seconds to process the information.

In 2026, when European leagues returned to empty stadiums, my model priced home advantage at 0.45 goals per match. After nine rounds without crowds, that figure fell to 0.08. Home is geography, until it disappears. I turned down a commission to write about crowdless football because I needed three more weeks of data. When the piece ran, I opened it by admitting that I had been wrong not to include the crowd variable in the first place.

That lesson carries over to the millimetre story. A forgotten variable does not collapse a model immediately. It makes the model right in a skewed way, and the skew only surfaces when an event large enough drags it above the waterline.

What is worth tracking in the next cycle is not camera resolution. It is the dictionary. When a stat graphic appears after a match and travels for two hours, the most useful question is not whether the number is right, but when that label was defined, by whom, and how many times it has changed since the season began.

"Getting one variable wrong is like losing your bearings for an entire year."

A season missing detail is like a match missing stoppage time: you still know the score, but you do not know what happened in the minutes nobody recorded.

"Data whispers. Whoever listens hears an entire match." But before listening, ask who taught it to speak.