When a Football Data Store Accepts a Record With No Football in It

There is a file sitting in our football analytics store with the word "football" in its first field. I opened it close to midnight, after rewatching an old match tape, and read seventeen information points from top to bottom. Not one sentence about football. No club. No player. No tactics, no contract, no transfer fee of any kind. Only a 38-year-old woman named Allison Holker talking about a gold chain and a set of clothes that still carry her late husband's scent — Stephen "tWitch" Boss, who died in December 2026 at the age of 40. Three children named with their ages: Maddox, 10; Weslie, 18; Zaia, 5. Two competitions appear: a television dance show from 2026 and a charity awards gala run by the dance industry, on 21 October. I read all of it. Then read it again. What is broken in my profession, it turns out, does not sit on the pitch.

Over the past decade, every article, every bulletin, every commentary segment after it goes to air has acquired a second life: it is fed into a content data store, where a classification layer assigns a subject label. That label decides where the record goes — a player statistics table, a public-sentiment heat index, or a transfer database. A mislabelled record does not make a sound. It simply sits there. And every time someone runs a query, it adds one more speck of dust to an index nobody audits.

I know this because I used to write the raw source material for those stores. In 2026, after the Beijing derby between Beijing Guoan and Shanghai SIPG at the Workers' Stadium, I stayed up all night writing an analysis of the 3-5-2 that coach Villas-Boas used for SIPG, showing how it dismantled Guoan's midfield. The piece was shared more than 2,000 times in 48 hours. What I did not know then: four years later, a paragraph of it would be read by a model and attached to some entity node, and nobody in the newsroom would be consulted.

A sound football data store has to survive two questions: who is this record about, and does it actually concern football. The second question is usually skipped.

The mechanics of this error deserve a close look, because it is a failure of an entire system layer rather than of one article. Seventeen information points, not one containing a football term. No xG, no PPDA, no formation diagram, no release clause. So where did the "football" label come from? Most likely a keyword-based classifier triggered by the surname "Boss" — a surname that appears densely in football at coach and director level. Or worse: a batch-level labelling error, in which an entire set of records was pushed into the same category.

The damage does not stop at one junk record. A mislabelled record pollutes the entity-linking dictionary, and that dictionary is what determines how every subsequent article is read. If the system has learned that "Boss" is a football name, then the next time a real coach named Boss appears, the model will be more confident than it should be — and confidently wrong.

What makes this case more uncomfortable than a single error is its periodicity. A memorial story has two fixed dates on the calendar: the date of death and the date of birth. Which means near-duplicate content will return to the data store every year, twice, indefinitely. An error left unfixed does not stay still — it reproduces.

And here is where the story touches real football. My industry calls pieces like this maintenance content: low emotional amplitude, high durability, no forward-looking claim, and therefore impossible to contradict. Football is full of them. Player birthday posts. Club foundation-anniversary statements. "Exactly ten years ago today" pieces. They carry no news, but they carry numbers. And in many internal dashboards, the volume of articles about a club is used as a proxy for public interest.

I have seen the consequences of that kind of measurement. In June 2026, in Kazan, when Kim Young-gwon scored in the 90+3rd minute and Son Heung-min made it 2-0 in the 90+6th, I sat in the commentary booth and said nothing for fifteen seconds. My editor later told me that silence carried more emotion than any line of commentary. But silence generates no data row. Measured purely by article count, that evening would score lower than a club anniversary statement published the same week.

In the source article, the sourcing chain is also worth studying. The exclusive interview belonged to PEOPLE, with a named subject and direct first-person quotes. The Express Tribune republished it and added no independent reporting. A source chain one layer deep. In football journalism, a source chain one layer deep is an everyday occurrence: one outlet runs a transfer story, ten others cite the first, and by the end of the day readers believe ten independent sources have confirmed it.

My defence against that is crude and manual: I log the provenance of every fact I use, with a date. I do it because I was once called "a commentator who only looks at players' faces", and I know what it feels like to be judged by a wrong label.

People say I only look at faces. Yes — because I believe the face is where the match deposits its pain. A player's face tells the story the scoreboard never shows. But reading faces is also a form of labelling: I see a frown and call it lost confidence; a data model sees the word "Boss" and calls it football. We commit the same error — reading too much from too little, and never recording our level of certainty.

The difference is that a writer can be challenged by readers directly beneath the piece, while a mislabelled record is challenged by nobody. It simply replicates itself in silence.

The first instinct in any newsroom is to blame the algorithm. I think that is a lazy reflex. A classifier that mislabels only mirrors the habits of the people who trained it: measuring presence instead of measuring value.

Try an uncomfortable comparison. Last season, how many articles about a mid-table club were produced by the club itself, and how many by people who actually had something to say? I do not have a certain answer. But I know which side wins every dashboard.

In 2026, when football returned after the shutdown, I went to Signal Iduna Park and recorded a full ninety minutes from an empty stand. Players said things to each other that spectators had never been allowed to hear. With no crowd, I still spoke. But I spoke so that the emptiness would know it, too, was a witness. Those ninety minutes produced no metric at all. They produced one article that left me exhausted, and perhaps that is why it was right.

In the coming weeks, more records will be loaded into that store, and a few of them will carry the wrong label. The work is to ask again the question every classification layer avoids: does this actually concern football? I will keep logging dates against every fact I use. I do it because I want the label above my own work to still be accurate when someone opens it ten years from now.