When UNAM Got Filed Under Football: One Mislabel and the Price Behind It

One record in the data system carries the label “Football”. Inside it are 24 information points: no team, no score, no contract, no player, no stadium. The only thing that looks like football is a four-letter string — UNAM.

The record's actual content belongs to a very different story. It is the march of 26 September 2026, when the parents and relatives of 43 disappeared students renewed their demand to know what happened, twelve years on. They protested outside the offices of Mexico's National Human Rights Commission (CNDH) and the Supreme Court (SCJN), met Chief Justice Hugo Aguilar Ortiz, requested the release of documentation relating to the armed forces, and awaited a government report on investigation lines scheduled for Monday, 28 September 2026.

Across all 24 information points, not one word belongs to football. And yet the system labelled it that way. That is where this story begins.

How such an error happens is no mystery. Content classification systems today run largely on automated entity recognition: they scan text, find familiar names, and infer a category. In a sports tagger's dictionary, “UNAM” is a real entity. Universidad Nacional Autónoma de México is Mexico's largest public university. Pumas UNAM is the club playing in Liga MX. Two entities with near-identical names. One line of copy mentioning that UNAM students joined the march was enough for the tagger to drag the whole record into the football drawer.

The rest of the text bears no resemblance to football at all. The headline, the lead, the quotes, every factual detail belongs to human rights and institutional accountability. A simple validation gate — does the text contain at least one genuine football entity, a competition name, a club, a player, a match — would have stopped the error before it reached the analysis queue.

In Vietnam, I have watched the sports content infrastructure long enough to know the same problem is always present. An aggregator page, an automatic translation tool, a transfer feed that works by grabbing a player's name and pairing it with a club: that is the ideal environment for a labelling error. The difference is that here the error rarely surfaces as a data label. It surfaces as a wrong headline that convinces thousands of people a player has left his club.

The greatest asset of a football data system is not the volume of records, it is the accuracy of the labels. Domain-label precision measures the share of records in a category that genuinely belong to it. A record labelled “Football” with no football inside is a noise sample. If it enters a training set, the model starts learning false signals — that a university's name indicates football, that protest vocabulary can appear under transfers. One mistake is easy to fix. A batch of them drags the precision of the whole data layer down, and reader trust down with it.

Rumours are not wrong — they simply arrive earlier than the truth. Bad data is different. It does not arrive early; it stays, because it lives on a hard drive rather than in memory.

One detail made me pause longer than the rest. Of the 24 information points, only three carry any source attribution: one cites a government announcement, one cites the families, one cites the Chief Justice. The other twenty-one are assertions nobody owns. No outlet, no byline, no wire.

I have seen this exact structure somewhere else: transfer feeds. A post says player X has joined club Y for a fee of Z billion dong. No source. No timestamp. No clause. Three days later, ten other pages repeat it. A week later, fans are fighting to defend a number that never existed.

Based on my experience watching matches and transfer windows, I always split a report into two separate questions. What does this data actually say? And who benefits when it spreads? For the mislabelled record, the first answer is “nothing about football”. The second is more interesting: nobody benefits. No deal, no negotiation, no winner. Just a machine running faster than its own capacity to check.

Mbappé taught me one thing: watching speed is fine, watching the direction of movement is better. Here, speed is the number of records the system processes each day. Direction of movement is where that record goes and what it is used for. A human rights record filed in the football drawer does no harm to football. It harms the story inside it: a real case, real people, real grief, converted into noise.

Writing for the forgotten is how I remind them I never stopped looking. A mislabelled record is a kind of forgotten thing too — forgotten by the very system that produced it.

The cost divides into three layers. Compute cost comes first: every bad record still consumes processing resources, and across a large batch that is real money. Model cost is heavier: each noise sample drags precision down, and falling precision means re-labelling from scratch — far more expensive than blocking the error at the gate. The hardest layer to repair is trust. When a football data product publishes a conclusion, the reader cannot verify it. They can only believe or disbelieve. Every public error makes the disbelief fraction a little larger.

The instinctive fix will be to re-tag, add keywords, tighten the filters. I think that fix is also how the real problem gets missed.

If an automated classifier cannot tell a football report from a human rights report, the question worth asking is not where the classifier broke, but why football coverage became so easy to confuse. Look at most of the football content produced daily — one-line transfer items, re-edited clips, scorer lists, template round-ups — and it has almost no domain signature left. The same sentence structure, the same headline pattern, the same habit of leading with a number. A machine reading thousands of those pieces will learn that “football” is a formatting template rather than a subject. And when the template is hollow, anything that accidentally fits slips through.

The second blind spot is faith in the process. Nobody caught this error at the editorial layer, because at the editorial layer there was nobody. It only surfaced when a person actually read the record and asked what was inside. The most reliable gate remains a human willing to read.

What I will be watching is not the mislabelled record itself, but the two things behind it. What the report scheduled for 28 September 2026 will say — a concrete checkpoint, with a date and someone accountable. And whether football data products build a validation gate exactly where one is needed. An empty stand does not mean nobody is listening, but a wrong label always means nobody read it properly.