A “Football” Label for a Horse Rescue in Tláhuac: A Classification Error and Its Data Cost

Santa Catarina highway, Tláhuac borough, on the south-eastern edge of Mexico City. A chestnut male horse, roughly eighteen months old, struck by a vehicle on the roadway. The Animal Surveillance Brigade (BVA) of the Mexico City Secretariat of Citizen Security (SSC) was dispatched to the scene, safeguarded the animal from moving traffic, and transferred it to the BVA facility in Xochimilco. Veterinary surgeons and zootechnicians would assess its condition there. The report stops at that turning point: a rescue, and an animal held for observation.

Then that report walked into a sports analytics pipeline carrying the label football.

Fifteen information points in the source. I read all fifteen. No team, no player, no coach, no competition, no governing body. The only entity set is BVA, SSC, Tláhuac borough, Santa Catarina highway, Xochimilco, and one male horse of roughly eighteen months. The xG shock at Hàng Đẫy turned me from a spectator into a reader of data, and since 2026 I have assumed by default that every error sits inside my own calculations. This time the error appeared before a single calculation was run.

A wrong label is cheaper than a wrong analysis

A sports data pipeline runs in three layers. Collection: news, press releases, wire copy, feeds. Labelling: topic, domain, article type, entities. Analysis: where I sit, where xG, PPDA, distance covered and contextual coefficients get computed. Public argument in this industry almost always aims at the third layer, because that is where predictions are produced. The second layer is the one that decides. A wrong label there travels into every model downstream, and no model detects it on its own.

I know this from the other side. In 2026, after Hà Nội FC drew 1-1 with Quảng Nam FC at Hàng Đẫy, I lost 180 million đồng trusting the shape of the game. Hà Nội took 17 shots for 2.87 xG; Quảng Nam took 2 shots for 0.94 xG. I sat down, reviewed 112 V-League matches from round 1 to round 14, hand-computed xG for every attempt, and found that Hà Nội's finishing ran 23% below the league average. A month later, their run of four straight defeats confirmed the table. The lesson I read then was: do not trust the eye, trust data you built yourself.

The second lesson took me years longer to read, and it came from an entirely different case.

Nine dimensions, eight empty cells

When I ran the fifteen information points of the Tláhuac report through the nine-dimension framework I use for matches, the result did not land in the “thin data” zone. It came back empty.

Tactical dimension: no formation, no system, no playing style. Financial dimension: not a single monetary figure — no fee, no wage, no valuation. League dimension: no competition is named, so food-chain position is undefined. Rules dimension: FIFA, UEFA and national associations are absent. Dressing-room dimension: no squad, no leadership group, no contract dynamic. Industry transmission: not one link of the football value chain is touched.

The only genuinely analytical finding sits in the source structure. Four information points are explicitly attributed to the SSC, the city government's security department. Nine carry no attribution at all. The whole narrative scaffold — headline, subheading, geographic framing, the causal claim “struck by a vehicle” — sits in the unattributed group. No veterinary clinic, no transport authority, no animal-welfare organisation was asked. The event is retold entirely through the lens of the agency that responded to it.

For a public-safety item, that structure is normal. The report is not at fault. The fault is that it was filed under football.

Why a wrong label travels far

A domain label is the cheapest thing to create in a data system and the most expensive to fix. It is cheap because one keyword match is enough. It is expensive because every layer behind it believes it.

An item labelled football will flow into fan-sentiment models, into keyword taxonomies, into entity-resolution graphs. In that last layer the concrete damage is this: the system will try to bind “Tláhuac”, “Santa Catarina” and “Xochimilco” into club and competition entity sets. Months later, when someone queries data on a football club, those place names can surface as noise, and nobody can trace them back.

The word “brigade” in Animal Surveillance Brigade is an obvious candidate. A keyword classifier can read that as an organisational-unit name and attach it to a sports entity set. That is a hypothesis about the pipeline, not about the report. A single case cannot establish the mechanism, but it is enough to raise an operational question.

Contrarian: the suspect sits on the label-consumer side

The first reflex is to blame the algorithm. That reflex is convenient, and it misses most of the problem.

The classifier is wrong because it was designed to be wrong in one specific direction: it prioritises coverage over precision, because missing a genuine football article costs far more commercially than mislabelling an unrelated one. Given base rates in news data, error in that direction is a rational output of a rational objective function.

The problem sits with those who consume the label. We treat a domain label as ground truth, as a verified fact, when it is only a probabilistic assignment. Belief is a noise variable; run the emotional regression before placing the bet. The same logic applies to belief in a pipeline: before using a dataset, run the test on its own labels.

There is also a storytelling temptation here. The romantic tale of “small data beating big data” sounds appealing and is wrong in exactly the same way as the tale of the small town beating the giant. No lone data point beats a system. What happens is that one wrong data point lands in the right place and breaks the system. Being 59 gives me the angle: every cycle is a loop with a remainder. The remainder here is a horse rescue, and it is recording our error, not the report's.

Correlation is not causation. An article landing in a category proves nothing about its content. It proves only that the category is being run on too broad a rule.

Signal for the next cycle

In my work, a broken model is a good day, provided I record where it broke. Kazan does not take revenge; Kazan simply keeps the table and waits for me to miscalculate. In 2026, before the World Cup group stage, I audited Germany's pressing data: average distance covered down 12.3% on the 2026 title-winning side, PPDA up from 8.2 to 11.7. I published a prediction that Germany would exit in the group stage. On the night of 27 June in Kazan, Germany lost 0-2 to South Korea with 0.41 xG. The table held, but only because I had checked its inputs first.

The Tláhuac case is a negative control. The correct output for this item is “no football content detected”. A mature pipeline must be able to produce that output and must record that it did. The cheapest fix is a content-validation gate ahead of the domain labeller, plus this item entered into the regression test set.

A broken model is the day the data monk has to burn his scripture back to the original text. The crowd leaves, the model breaks, and I learn to hear the breathing of the empty stand. This time the empty stand sits on a motorway on the edge of Mexico City, where an eighteen-month-old horse is being held for observation, and one data layer far away from it, someone is calling that football.

Behind the database running at your back, how many other labels are wrong in exactly this way — and will you find them before or after they have entered a decision?