Domain Error: How a Tepotzotlán Crime Report Landed in the Football Pipeline

Five minutes and fifty-one seconds of CCTV footage. I have stood on the touchline for twenty-eight years, and in that time I learned to read three documents — the scoresheet, the team sheet and the deal sheet. All three share one thing: date first, then name, then authority. The paper that landed in front of me this time was none of those. It was a metadata field. It said: Domain Label: football.

Inside the footage was something else entirely. A street in the Loma Bonita colonia of Tepotzotlán, State of Mexico; an elderly woman; two people on a motorcycle; an attempted snatch of her bag; the bike going down; neighbours gathering. The state prosecutor's office, FGJEM, later confirmed two detainees — Laura 'N' and Argelia 'N' — and referral to the Ministerio Público.

There is not one club in this report. Not one player. Not one coach, competition, transfer, wage structure or governing body. And yet the file arrived in the football analysis pipeline. This article is about that event — the dataset event, where seventeen information points yielded zero football entities.

The pipeline runs in two stages. Stage-1 deconstruction extracts facts, claims, entities and source quality, then attaches a domain label. Stage-2 analysis takes that label at face value and stacks tactical, financial, governance and risk layers on top. The label is not an opinion; it is the latch on the door. Fit the latch in the wrong place and everything that walks through sits down in the wrong room.

FGJEM and the municipal government of Tepotzotlán are both institutional sources and reliable for procedural facts. The CCTV's provenance is weak: the video is said to have spread on social media, with no chain of custody, no named outlet and no byline. And the date is only this — "Thursday, October 29." No year. October 29 fell on a Thursday in 2026, in 2026, and will again in 2026.

I learned to read the option before I read the headline; now I read the field before the headline. And this field was lying. The question is not how the error happened. The question is why nobody caught it, and how much damage it could have done.

The likeliest source is keyword matching. Spanish crime vocabulary and football vocabulary share verbs that look identical in a dictionary. Interceptar — to stop a pedestrian, and to intercept a pass. Marcar — to mark a target in a police operation, to mark a player on the pitch. Barrida — a sweep in an operation, a slide tackle in a match. From forcejear to detener, every word has two faces. An automated rule built to spot football terminology will wave these through.

The second likely source is the loss of section metadata at feed level. On a news site this story probably sat on a general-news page, with a 'Deportes' section link or ad in the margin. Extraction collapses page structure, and the tag that survives belongs to the layout, not the content. What is most striking here is that across seventeen information points there is not a single partial football signal — no club, no city-club, no league, no stadium. That is exactly what makes it an ideal negative control.

In a lab you need a sample that is definitely not what it claims to be. This file is that sample. The real test of any classifier is the item where the overlap between claim and content is zero. Here the overlap is not zero but negative — the content belongs to another profession, another country's institutions, another legal framework.

So the thing that matters is not the analysis. It is the gate. The cheapest and most effective gate is an entity check: no item qualifies for the 'football' label unless at least one club, player or competition is present. That check belongs before the football taxonomy, not after it. The room where deals are briefed is never the room where they are done — and here the briefing room was at the wrong address to begin with.

The date problem is this file's second great weakness, and it is a familiar one in my trade. "Thursday, October 29" is a timestamp that fits three different years. In transfer windows I have seen such claims hundreds of times — "medical on Monday" — when the medical date, the announcement date and the registration date are three separate things. A deal sheet without a year is not a deal sheet. It is a rumour that has filled in a form.

Source tiering splits this file in two. The institutional half — the FGJEM statement, the municipal statement, PDI confirmation of investigation — is solid. The video half is weak: "spread on social media" means nobody knows when the clip was filmed, who filmed it, or what was cut. The five-minute-fifty-one-second runtime suggests a compilation from multiple cameras rather than a single capture — standard practice for municipal surveillance releases. The chain of custody is still zero.

This is where an old habit earns its keep. Rumours come in three layers: the "sources understand" layer, the "it is believed" layer, and the document layer. I write only from the last. This file had the first two and lacked the third — except that the FGJEM statement is an institutional document. So the facts of the incident cannot be questioned. The label someone glued on top of that statement can, and should, be.

The real risk is plain. Hand this input to any model or analyst labelled 'football' and it will answer, because it is asked to answer. It will invent PPDA, invent xG, invent a formation, invent pressing triggers — and it will read so smoothly that no reader will doubt it. From years of watching matches I know one thing clearly: a wrong number is more dangerous than a wrong analysis, because nobody argues with a wrong number. They cite it.

It reminds me of one of my own habits. In 2026, with stadiums empty and club revenues collapsing, I started building a database of loan-with-option deals and abandoned it twice. The same question that applied to that database applies here: did you check every row, or did you trust the column names? A column name is a label. Labels sometimes lie.

One more thing hides here, and pipeline engineers routinely ignore it — editorial convention. The source names the suspects Laura 'N' and Argelia 'N' and uses the language of 'probable participación'. That is the standard Mexican journalistic practice of presumption of innocence. If this item is ever repurposed into another product, that convention travels with it. Turning anyone into 'the accused' and adding full names would be a different kind of error — the kind that comes not from a lawsuit but from editorial carelessness.

Now the uncomfortable part that gets said least. Football content pipelines have invested everything in ingestion — more feeds, more outlets, more items. Nothing was invested in the gate. Plenty of searching, not one latch. And with no latch, a dataset slowly fills with claims that have no documentary basis at all.

The second counter-intuitive point: the most dangerous item is never the one that is a hundred per cent wrong. That gets caught immediately. The dangerous one is eighty per cent right — right league, right club, right player, wrong wage band. The Tepotzotlán file was a hundred per cent wrong, which is why it surfaced. The file that never surfaces keeps circulating like a communiqué.

And the third point, for anyone treating this as an isolated incident: the collision between football vocabulary and crime vocabulary is not a Spanish problem. In Hindi and English too — to mark, to intercept, to block, to bounce — these are simultaneously sporting verbs and administrative ones. The day a classifier relies on them alone is the day it produces a hundred such errors.

So what next? My recommendation is an audit, and one with a timestamp. Pull a fixed sample from every thousand items in any corpus and ask a single question: does this item contain at least one entity matching its label? If the error rate clears two per cent, the problem is not the keywords — it is the missing gate, and the whole batch should be quarantined.

Follow the clause and you will find the club — that is my old rule. This file had no clause, only a metadata field. The rule held anyway: I followed the field and found the pipeline. The question is how many items in yours would fail their own label check right now — and whether you have set a date to run it.