A Football Label on a Fashion Interview: Audit Notes on a Pipeline Classification Failure

On the night of 12 February 2026, I opened a record in my data pipeline. Classification label: football. Fifteen information points, neatly numbered one through fifteen, fully structured like every other record that day.

I counted. Clubs: none. Players: none. Coaches: none. Leagues, matchdays, contracts, transfer fees, xG, PPDA, league tables: none. What I held was a magazine interview with a 60-year-old former supermodel, discussing a decision to pose for a men's magazine decades earlier, a daughter who followed her into modelling, and a marriage being redefined after the children left home.

Football content inside a record labelled football: 0 out of 15.

Nobody posts this kind of error on social media. It does not create a wrong goal, does not distort a league table, does not cost anyone money. It sits quietly in the pipeline, waiting to be folded into an aggregate index, waiting to be fed into a model trying to estimate something about Vietnamese football.

I stayed up until three in the morning to write the audit note. There is nothing wrong with the interview itself. It is wrong in a different place: it was filed in a drawer that does not belong to it.

A pipeline does not know it is wrong

Every day, sports content aggregation systems in Vietnam swallow thousands of articles. Each article passes through a classifier and receives a label: V.League 1, transfers, national team, international football, or an unrelated drawer. The label decides where that article goes — into the news feed, into the training stream for a model, into a transfer-market tracker, or into the bin.

In seven years of working with football data, I learned something no classroom teaches: error at the ingestion layer is multiplicative, not additive. One bad article inside a set of 500 gives a contamination rate of 0.2 percent, which sounds harmless. But if that rate holds at two percent, then one in every fifty articles carries the wrong label. For a system processing 200 articles a day, that is four contaminated items daily, 120 a month, nearly 1,500 a year.

I have a personal example of how the ingestion layer determines the conclusion. In 2026, when V.League played in empty stadiums, I analysed 156 matches and found the home-win rate fell from 46 percent to 38 percent. I published that result; an analyst at a club used it to adjust how they weighted away fixtures.

Empty stadiums do not erase the truth. They only strip away the fog that 40,000 voices once created.

But if three of those 156 matches had carried the wrong label, my conclusion would have collapsed. I would not have dared to publish. I would have sat counting another season and missed the window when the information still mattered.

Trust in data does not come from the model. It comes from knowing exactly what got in.

The audit note

This is the form I apply to any anomalous record, wherever it comes from.

| Check item | Result | Acceptance threshold | Status | |---|---|---|---| | Football entities in the record | 0 of 15 | at least 1 | Fail | | Information points with independent sourcing | 3 of 15 | at least 10 | Fail | | Publication date | Absent | Mandatory | Fail | | Byline | Absent | Mandatory | Fail | | Internal dating marker | Contradictory | Must reconcile | Fail | | Error rate in the same stream (sample of 200 records) | 1.5 percent | under 2 percent | Conditional pass |

Three rows deserve a pause.

The first: twelve of fifteen information points are sourced as "none", or sourced to the subject of the piece herself. Three quarters of the raw data in this record is self-reported, with no verifier and no cross-document. In my workflow, an information point without independent sourcing is parked, not modelled.

The second: no byline, no publication date, no issue number. The distribution source is an English-language newspaper republishing the content. A record with no timestamp cannot be assigned to any data series, because I do not know which season it belongs to.

The third: the only dating clue is the age figure — 60. Checked against the subject's public date of birth, that number pushes the article outside the pipeline's time window. I marked it "data to be verified" and moved on. A record that contradicts itself on time is a record that cannot be trusted on any other axis.

My hypothesis for the cause: the classifier was defeated by surface-level lexical overlap. The piece contains phrases belonging to sports vocabulary, and a magazine name previously associated with sports interviews. The classifier does not read meaning. It counts words. And it counted correctly in the wrong way.

I tested a second hypothesis: whether this was an isolated failure. I sampled 200 random records from the same stream and found two similar cases, both entertainment content labelled as sport. A rate of 1.5 percent. Below the two-percent threshold I set for myself, but enough to open a standing audit ticket.

A single number can lie, but a model validated across 10,000 matches has no reason to pretend. The problem is that this model has never seen what I just saw.

What a negative control is actually worth

In statistics, a negative control is used to test whether a system raises false alarms. You feed it something guaranteed to contain none of the target, then watch whether it labels the target anyway.

This record is a perfect negative control. I know with certainty it contains not one football unit. Yet it sat in the football drawer.

For a newsroom, its value lies elsewhere. An article about a supermodel is not a threat to a V.League feed. The threat is habit: we build a filter, we trust the filter, and we never check its output by hand.

I have watched this across the industry. V.League clubs have started using player-tracking data, buying metric packages, hiring analysts. Vietnamese sports newsrooms have started building reader-facing data tables. Aggregation platforms such as VuaBong.vn handle far more content volume than they did five years ago.

But the ingestion layer is still one person with a spreadsheet and a belief that every source is equal. A repost from a foreign newspaper is treated the same as an official release from a competition organiser. An undated interview is treated like a match report with a paper trail.

The gap between those two things is the entire value of data journalism. If I treat them as one, I have nothing left to sell but my own name.

The counterintuitive angle

When the press room laughed at xG, I knew I was reading exactly the book they had not opened.

But the lesson from this record does not sit with the sceptics. It sits with the believers — myself included.

Vietnamese sports data is spending most of its energy on model questions: xG or shot counts, how much pressing weight is enough, regression or neural networks. Every one of those questions is misplaced if the input data has not been verified.

A good model running on dirty data produces a wrong conclusion with confidence. A mediocre model running on clean data still produces a usable conclusion. In seven years I have never seen anyone promoted for personally auditing the provenance of a dataset. I have seen plenty promoted for building an elegant model on a dataset nobody ever checked.

The most counterintuitive part: articles like that supermodel interview do not damage the pipeline because of their content. They damage it because they are never rejected. A pipeline only stays clean if it can say "no" to an article that sounds plausible. That capability has to be built by hand, with time, with a person accountable for it.

In Vietnam, that infrastructure is being built more slowly than the models are arriving. We import the model first, then go looking for rules for the data.

Signals for the next cycle

If you build a sports data product in Vietnam, here are three things you can start this week.

Every week, pull fifty records at random that have already passed the classifier and read the labels by hand. Log the error rate. If it exceeds two percent, stop and inspect the filter before doing anything else.

For every republished article, record three things: the original source, the publication date, the byline. If one is missing, the article does not enter the model — it enters the holding queue.

Build a fixed negative-control set, one hundred records guaranteed to be unrelated to football, and run them through the classifier every time you change the code. It is the cheapest way to learn what you just broke.

I am not proposing anything new. I am proposing something very old: count again.

Based on my experience tracking matches and data streams, I expect that within twelve months the ingestion error rate will become the third most requested metric in Vietnamese sports analytics rooms, behind expected goals and sprint distance. Whoever builds the audit threshold first keeps readers longer. And whoever keeps readers longest will be the one willing to tell a very plausible article: no.