A Power-Outage Notice From Torreón Landed in a Football Data Lake: One Labelling Error and Its Price

October 17, 2026 falls on a Saturday. In Torreón, Coahuila, Mexico, the Ampliación Lázaro Cárdenas neighbourhood and the streets Héroes de Río Blanco, Lorenzo Ávalos and Tampico will lose power from 10:30 to 18:30 so that CFE — Comisión Federal de Electricidad, Mexico's state-owned utility — can carry out grid maintenance. The same day, the town of Motul in Yucatán goes dark from 07:00 to 14:00. Amatenango de la Frontera, Chiapas, from 07:00 to 15:00.

Nineteen information points. Not one club. Not one player. Not one match. Not one stadium. Not one transfer contract.

The record entered our analysis system tagged with a domain label: football.

I read it three times. The first time, to check whether I had missed a sentence mentioning Santos Laguna or Venados FC. The second time, to check whether the record had been truncated. The third time, I sat still, because I realised something worse than a wrong article: a machine was already primed to keep writing.

Our pipeline runs in two decoupled steps. Step one reads the text, extracts the information points, identifies the source's stance and purpose. Step two assigns a domain label, and that label decides which analytical framework gets invoked. With a football label, the system opens nine dimensions: tactics and technique, club finance and the transfer market, results and the opinion cycle, league landscape and team positioning, rules and governance, management and dressing room, risk profile, media narrative and expectations, and finally industry transmission.

Those nine dimensions are heavy machinery. They are built to answer specific questions: is the pressing system sustainable, does the wage bill breach a threshold, is a three-match losing run a long-term signal. Feed them the wrong raw material and they do not complain. They keep running.

The first xG table I ever wrote was by hand on a bus, back when nobody called it data yet. I sat in the back row and pencilled every move from fourteen V.League clubs in the 2026 season into a squared notebook, because I believed a shot from the edge of the box and a shot from the penalty spot could not be counted the same way. That belief was right. But what I learned afterwards is what kept me in this trade: data is only honest when the label stuck on it is honest too.

A record can be right in every detail and still be entirely worthless, simply because its label is wrong by one word.

Looking at the content, this record is of higher quality than many football analyses I have read. Nineteen information points, most of them traceable directly to CFE as the primary source: which body issued the notice, which maintenance work, which time windows, which areas, what scope of impact. Only the household precautions were added by the outlet — advising residents not to open the refrigerator repeatedly so food keeps, and to unplug sensitive electronics before power returns because voltage spikes on re-energisation can damage them. That is the only part without CFE behind it, and also the only part whose credibility cannot be graded, because the record stores neither the outlet's name nor a byline.

Put differently: the extraction step worked. The labelling step failed. The two never looked at each other.

The extraction step and the labelling step are running independently, and the gap between them is exactly where dirty data is born.

The cheapest guard I know is a single entity gate placed before the framework opens: require at least one football entity — a named club, player, coach, competition or stadium. This record has zero. The gate closes. The running cost is close to nothing, compared with the cost of a fabricated analysis published under a data journalist's name.

What chills me is what would happen if someone forced this record through all nine dimensions.

The tactical dimension has no formation, no system, no passage of play. The machinery would have to invent a story. And it would invent one very well, something like: an eight-hour daytime outage window reflects a congested fixture calendar, forcing the operator to compress maintenance into a midweek slot. It reads smoothly. Not one word of it is true, and not one word of it is grammatically wrong.

The finance dimension would conjure a wage bill, a broadcast revenue structure, a net debt ratio. The results dimension would record recent form on a sample of zero matches. The governance dimension would test FFP and PSR against a subject that does not exist. The dressing-room dimension would hunt for age curves and contract years among people who are never named. The transmission dimension would try to wire academies, clubs, broadcasting rights and derivative markets to a domestic power-outage notice.

The correct output of all seven of those dimensions is null. Not null out of cowardice, but null because there is no object to analyse.

One detail in the record is the easiest place to slip, and I want to linger there longer than anywhere else. All three affected states sit on Mexico's professional football map. Torreón is the home city of Santos Laguna, a Liga MX club. Mérida, capital of Yucatán, is home to Venados FC of Liga de Expansión MX, while Motul is merely a small town in the same state. Chiapas has had no stable top-flight football since Jaguares and Cafetaleros left.

A hurried writer would stitch those three facts into a headline. And where exactly would that headline be wrong?

Wrong because the Torreón outage zone is a residential district, not a stadium district. Wrong because the record names no venue, no match, no kick-off time, no broadcast window. Wrong because no path runs from a domestic power notice to the transfer market, to a wage bill, to broadcast revenue, without exactly the kind of data this record does not carry.

Geographic coincidence is the raw material of rumour, not the evidence of impact.

I have stood on the other side of this test. At the 2026 World Cup I built a PPDA model to measure pressing intensity and stopped for a long time on Croatia. Before the Argentina match, Croatia under Zlatko Dalić pressed frontally at a PPDA of 7.9, lower than sides famed for possession control. I wrote a long piece predicting they would reach the final and was laughed at for rating Croatia when nobody else did. The world saw Croatia as an underdog; I saw them as a coefficient chain nobody had dared to mine.

But let me state clearly what few people say after such a prediction lands. It landed because the coefficients were labelled correctly. If someone had mislabelled PPDA 7.9 as average possession minutes, I would have read Croatia as a slow, deep-defending side and written the exact opposite — in the same confident voice, with the same table, in the same argumentative structure. No reader could have told the two pieces apart on style alone.

Back to V.League. In 2026 my fourteen-club xG model had one bad cell. A header was logged at 0.4 instead of 0.04. For the next two rounds, a centre-back topped the league's xG/90 table. Nobody caught it, because the table still looked tidy, still matched the format, still printed, still got shared. My model does not cry and does not celebrate, but after every match it owes me a lesson. That day's lesson: an error at cell level propagates to table level; an error at label level propagates to conclusion level. Same mechanism, different scale.

Then came 2026. Empty stadiums, suspended leagues, no matches to analyse. I spent six months mining V.League data from 2026 to 2026 and found a pattern: clubs that changed chairman mid-season saw their win rate fall by 23 percent over the following five matches, driven by governance disruption. One club executive called to thank me for helping them delay a managerial sacking at a sensitive moment.

That 23 percent figure only stands because every record in the dataset carried its correct label: this is a match, this is a club, this is a season, this is a chairman change. If an administrative notice had slipped into that set, the rate would skew, and worse, it would skew in a way nobody could trace back to source. In 2026 the stadiums were empty, but every ball still landed in a model cell, and I understood that data never keeps company with a pandemic.

The risk matrix of this record reads strangely. Sporting risk: null. Financial risk: null. Personnel risk: null. Rules risk: null. Reputational risk: medium, and only in the sense of electricity customers complaining if power does not return on schedule — a service matter, not a football one.

Systemic risk, though, is high, and it has already materialised. A non-football record travelled through a football analysis pipeline carrying a football label. If it were published, the damage would not sit in an article about Mexico. The damage would sit in the fact that my readers would have one more reason not to trust any table I ever hand them again.

The threshold I set myself: two mislabelled records in one batch means suspending the pipeline and re-validating the whole thing. A single error is an incident. A cluster is a disease.

There is one small detail in the record I want to keep as a professional memento. October 17, 2026 is indeed a Saturday. Among nineteen information points, that is the only claim verifiable by calendar arithmetic, and it checks out. A nineteen-point record, and exactly one point measurable by mathematics.

The standing fear in my trade is that machines make things up. The real fear is different and far more uncomfortable: machines make things up fluently. Correct format. Correct terminology. Correct argumentative structure. Correct even in the modest tone of someone who knows a model's limits. Forced through a football framework, this record would produce phrases like fixture congestion, administrative stability, maintenance cycles — phrases I myself have used hundreds of times. And precisely because I have used them hundreds of times, no reader would have any way to detect that they were being used to describe a power cut.

The paradox is that we spend money on bigger models, denser tracking data, more advanced metrics, while the thing that stops this disaster is a single entity filter and one person reading the record again. Fancy tools do not rescue a wrong label. Writing by hand is not obsolete, and I do not say that out of nostalgia.

There is another reading, and I think it is more useful than criticising the system. This record is a perfect negative control. Any future version of the pipeline that can produce a tactical analysis or a financial table from this input has just proven, by its own output, that it is hallucinating. I am keeping it in the regression test set. A CFE power-outage notice dated October 17, 2026 is the cheapest yardstick I own for checking whether my model is still honest.

This transfer window, we are teaching readers a good habit: rank every rumour by evidence, separate noise from signal, ask where the money is, ask how the release clause is written, ask what the agent gets. The transfer market is a game for those who see far, not those who see much; value always arrives after patience. But the same habit must be applied to invisible records too — the data rows nobody has read, nobody has labelled, nobody has signed.

The lesson is not in Mexico. It is in the fact that we are building ever-larger football data lakes, ever-faster, while the labelling step remains the least scrutinised and least owned step of all. A wrong label does not ruin one article. It ruins the reader's right to trust every other article that follows.

Spectators watch the passage of play; I watch twenty-two numbers in motion — and wait patiently for them to tell a different story. But before waiting for numbers to tell stories, I have to be sure I have stuck the right label on the box that holds them. The question I carry into the next data batch is simple, and I want it to become the question of the whole trade: who signs their name to your label?