A Loaf of Bread in the Football Database: On Analysis Without Provenance

23:40. I open a file tagged “football”. Inside there are 25 information points. I read all of them, then read them again. No team. No player. No scoreline, no line-up, not a single minute of any match. The only thing that surfaces is a round loaf of bread dusted with sugar, decorated with strips of dough arranged in the shape of bones, eaten on 1 and 2 November every year in Mexico.

My silence did not come from surprise. Twenty years around football have taught me that mistakes are routine, and most mistakes do no harm because somebody sees them. I was silent because I knew exactly what would happen next. The file would enter an automated system. That system would be instructed to analyse “this football article” across nine dimensions: tactics and technique, club finance, the transfer market, league standing and team positioning, rules and governance, the dressing room, risk profile, media narrative and expectation, and industry transmission. And it would answer. Fully. Fluently. Not one line true.

That moment taught me something I had never written down: the problem with this trade is not that the machine writes badly. It is that the machine writes far too well about something that never existed.

At La Commanderie, the morning session begins before there is enough light to see the touchline markings. I arrive early as always and stand at the eastern fence, where I can hear the ball bouncing on grass without anyone noticing me. Forty metres away is the analysis room, where a computer runs all day. It runs while the squad trains, while the press conference happens, while I sit writing this.

My trade now has three speeds. The speed of the training session: slow, repetitive, mostly meaningless, occasionally exposing one detail worth writing. The speed of the newsroom: fast, measured in minutes, always asking for more. And the speed of the data pipeline: instantaneous, unwatched, unchecked.

The press-room door closes and I start to hear the match more clearly. I wrote that sentence years ago and still find it true. But it now needs a second clause: when the data door opens, I start to hear matches that never took place.

How a file goes astray is simple. The system captures text, assigns a domain label, stores it, then passes it to an analysis model. At the labelling stage, one field is filled in wrongly. I can count three plausible explanations. The label was inherited from a batch job mixing several subjects. Or a classifier tripped on a false-positive keyword somewhere on the page, a navigation tag reading “sport/culture”, say. Or the label field was left empty and the system defaulted to its most common label.

All three belong to operations, not to intelligence. And none of them was blocked, because no validation gate asked a single question: does this file contain at least one entity belonging to the domain just assigned? For the 25 information points in the bread file, the answer is no. Not one football entity. Yet the label sat there, ready to travel onward, with everything behind it built on the assumption that the label was right.

I call this a positional error. The content can be accurate, the prose can be good, the writer can mean well, but shelved in the wrong place it becomes a false source. A data label travels further than a text, because a text is read once while a label is copied without limit. An article about bread sitting in a culture archive is ordinary. The same article sitting in a football archive is an incident, because every subsequent query will pull it out and every subsequent model will believe it.

What troubles me is less the false content than how the system will respond to it. A model instructed to fill nine dimensions will fill nine dimensions. It has no internal incentive to say “I do not know”. The completeness of the template is itself the attack surface. And if nobody reads the output, that completeness drifts into the knowledge base as fact.

Across the 25 information points, not one leads to a named institution. No ethnographer, no cultural body, no university, no interviewee. Every source field reads the same handful of phrases: none, popular interpretation, popular stories, article author. No byline. No outlet. No publication date. The title asserts the year 2026.

I do not rush to conclude. I have been in this trade long enough to know that a future-year title on undated text can signal search-optimised content, and can equally be an outlet publishing ahead of the festive cycle. Both are plausible. What is not plausible is that a file with no date, no name and no source is allowed into a database used to generate analysis.

One detail stops me filing this under rubbish. The text itself refutes a popular legend about the bread's origin, arguing there is not enough historical evidence to affirm that origin, and it downgrades symbolic readings to “popular interpretation” rather than settled fact. That is claim discipline of a decent standard, better than a great deal of football copy I read each morning. But claim discipline is not sourcing. An article can be scrupulously cautious and still untraceable. The two axes are independent, and the bread file scores high on the first and low on the second.

In 2026, in Doha, I took a call at 2 a.m. The Olympique de Marseille communications manager told me Florian Thauvin, number 26, had pulled a hamstring in a closed session. That is the kind of item you publish and watch the traffic arrive. I did not publish.

I spent the next two hours calling the fitness coach, calling the club doctor, then calling Thauvin himself to ask how it felt, not when he would be back. The piece I wrote was headlined “The fear is not in the pain”, mentioned no recovery timeline, and dealt only with a player's fear of being forgotten. It set my monthly reading record. What I kept was not the record. What I kept was those two hours. An article without sources has no shelf life; it holds only until somebody checks it.

In 2026, my first assignment on the Olympique de Marseille beat was Ligue 1 matchday 12, a 1-3 defeat to Lyon at the Vélodrome, which holds 67,394. Afterwards a veteran reporter blocked me at the door: “The dressing room is not for girls.” I did not argue. I stood in the corridor for four hours and took notes. I watched Morgan Sanson, number 8, withdrawn on 62 minutes, walk past without looking at anyone, and I understood his frustration from the way he kicked a water bottle. The piece “Those who are not allowed to speak” came out of those four hours, and the head coach at the time shared it on his own page.

Moscow taught me that the truest sources rarely carry a business card. In 2026, aged 28, I was the only woman among the 12 reporters travelling with the France squad, and I was seated in an area with no wifi. I got talking to Ivan, the postman at the Saint-Denis training ground. He told me players hid sweets in their jacket pockets when leaving the hotel. When France beat Belgium 1-0 in the semi-final, I was the only one who wrote about Paul Pogba calling his daughter after the match. Nobody else knew, because nobody sat at the post desk.

I tell these three stories to set two ledgers side by side. On one side: four hours in a corridor, two hours on the phone, one postman, one water bottle. On the other: 25 information points with no source, no name and no date, believed absolutely by a machine because of one field.

I hold a professional bias I do not hide: the heat map has become football's new form of divination. A dark patch of colour does not tell you whether the player ran there because the system made him run there or because he did not know where to run. It tells you only that somebody ran. But heat maps are pretty, legible, shareable, and so they replace watching the footage back.

The bread file is the extreme form of the same disease. There, a label was assigned with no witness. Here, a patch of colour is read with no context. Both are data without provenance, and both are consumed the same way: fast, tidy, unverified.

The same thing happens to the tactical stories I read every week. When a big club goes out of a domestic cup, the familiar template appears within ten minutes: miracle, shock, spirit. I have watched enough knockout ties to know that most of those shocks are the inevitable consequence of two specific decisions: the stronger side rotates and drops its pressing intensity, while the weaker side pushes a high press into exactly the flank the stronger side has just substituted. That is a technical story, measurable through the passes allowed per defensive action, the PPDA metric the analytics community has used widely since Opta formalised it. But the technical story takes two days. The miracle story takes two minutes.

The return of the back three is the same pattern. I do not believe it is tactical progress. Most of the shifts to three centre-backs I have watched in Ligue 1 are reputation-defence reactions: the back four has just been broken and the coach needs a shape that looks like a solution. Nobody writes that headline, because it does not sell. A template filled with “tactical flexibility” sells.

The reflex is to blame artificial intelligence. I find that framing convenient and misdirected. The machine in this story did exactly one thing: it complied. It was told to analyse a football article, and it analysed one. The risk was created by a process that handed out the task without checking the input, and above that by our own habit of treating completeness as a mark of quality.

The frightening part is not that the machine writes wrongly. It is that it writes wrongly in a tidy, structured, apparently cautious way, and nobody has any incentive to read it back. A chaotic analysis gets stopped in thirty seconds. A fluent analysis of a loaf of bread can sit in a database for years, waiting to be cited.

The irony is that the cost of fixing this is close to zero. A single validation rule — does the file contain at least one entity belonging to the assigned domain — blocks it with perfect precision. One mandatory field for outlet, byline and publication date blocks most of the rest. We are paying an unlimited price for a fault that one line of configuration would patch. The reason we do not patch it is that the people who bear the consequences are not the people who run the pipeline.

This season, when I read a report with no byline, no outlet and no date, I will ask myself who verified it and with what. Not to doubt the writer. To know whether I am reading an observation or reading a data label.

Every season is a heartbeat, and I am only trying to catch the right beat. That heartbeat now has an extra step nobody teaches in journalism school: