An Obituary Filed Under Football: The Failure Is in the Label, Not the Text

3:47 in the morning. Manchester rain, the November kind that never quite stops. I opened my scouting database — the sediment pit, as I call it, where every record is a layer of soil — and found a file with a familiar label: football.

Inside was the obituary of Tom Skerritt, the American actor, dead at 93, recorded as natural causes. Twenty-five information points. I read all twenty-five, slowly, the way I read a scouting report.

No club. No player. No coach, no competition, no transfer, not one line on wage bills or financial fair play. There was Alien, there was Top Gun, there was MASH, there was Picket Fences. There was a Primetime Emmy in 2026, a nomination in 2026, several Screen Actors Guild nods. There was Wayne State, there was UCLA, there was a spell of service in the United States Air Force. There was a wedding date. There were surviving family members.

And above all of it, one label: football. I sat still for three minutes. Not out of surprise. Because I realised I had once come close to believing a record like this, simply because it sat in the right place.

What I sell is methodology, not players. In March 2026, when the Premier League suspended and my freelance contract with The Athletic was cut, I spent six football-free months building my own scoring system, the Youth Impact Index: ten criteria, measured across three consecutive seasons to strip out single-season noise. When football returned in June, academies had lost almost all their data because youth competitions had been cancelled. Huddersfield Town paid 15,000 pounds for a report on five Brentford youngsters. Since then, clients buy confidence in a process, not confidence in a name.

My view of the transfer market has not moved in sixteen years of watching it: the race between the giants is a brand arms race, and the genuinely valuable contract sits at small clubs — where a 15,000-pound report can turn a season. Which is exactly why small clubs suffer most from dirty data: they have no budget to hire someone to re-check the source.

Every report starts with a raw record. Raw records arrive from aggregators: something scans the headline, scans the metadata, assigns a domain label. Nobody in that chain looks at the actual content.

In June 2026 I was sent to Moscow. On 30 June I stood in the Luzhniki corridor after France against Argentina, listening to two German scouts talk about Kylian Mbappe, then 19: quick, but cannot sustain it across ninety minutes. I wrote a 2,000-word rebuttal, not to defend Mbappe but to point out that they had concluded from too short a sample. An editor at The Athletic noticed it. The call of 2026 saved nobody's career, but it saved me from my own arrogance.

The lesson I kept was not about Mbappe. It was that a conclusion is only as reliable as the layer of data that produced it.

Four layers to this record.

The entity layer. Across twenty-five information points there is not one football entity. The names cited — Skerritt, Sigourney Weaver, Tom Cruise — are performers. The works cited — Alien, Top Gun, Picket Fences, MASH — are films and television series. Ridley Scott and Robert Altman are directors, a category with no legitimate mapping onto tactical analysis. At this layer the answer is null, and null is the only honest answer.

The labelling layer. The record is complete at information-point level: twenty-five points, coherent, each tied to a paragraph and a source. The deconstruction ran correctly. The error sits one layer higher, at domain assignment. Most likely the label came from a headline or a metadata field rather than from entity inspection. If so, the system has no post-labelling validation gate at all.

The date layer. The report dates the death to Wednesday, 21 October 2026. I checked a calendar: it is indeed a Wednesday. That detail is what made my skin crawl. A typo normally breaks the weekday-date check and we catch it immediately. Here the check passed. The discrepancy lives in the publication-date field or in the source, not in a typist's finger.

The confirmation layer. The chain is two steps: a publicist told Variety, Variety was relayed by The Express Tribune. No family statement, no independent confirmation. In football this structure is utterly familiar: an agent tells a trade outlet, the trade outlet is relayed by a daily, and we call it a second-tier source and mark the reliability down. With an obituary we forget to perform that move, because grief has no price-inflation motive.

A mislabelled record does not ruin a match. It ruins a model, and that model will pass judgement on a seventeen-year-old.

The propagation mechanism is uncomfortably simple. A contaminated record enters the database carrying a vocabulary cluster around death, memorial, bereavement. Six months later, when a former player dies and someone queries the archive, the system returns a result set with degraded precision. Worse: if a model trains on that archive, it learns the weights of an irrelevant vocabulary and then applies those weights to reports about young players. Nobody notices, because the output still reads smoothly, still holds its format, still looks like a decent report.

In September 2026 I was an analysis assistant at the Manchester City academy, assigned to track Phil Foden, then 16, in an U19 training match. I wrote twelve pages concluding that the boy lacked the pace and the frame for elite football. Three months later Foden was promoted and scored on his Champions League debut. I was wrong because I read the physical data and ignored how he read the game. Based on my experience watching matches across English academies since 2026, what gets missed is almost always what does not sit in a numeric column.

Before I write a star's name, I have to peel off a thick layer of soil called aura. But there is another layer just as thick and far less discussed: the layer of the label. A wrong report is like a broken shard of pottery: handled carelessly, it cuts the hand of the person who wrote it.

The instinctive reaction will be to blame the classifier. Swap the model, add keywords, done. I do not think that is where the repair matters most.

The classifier is a symptom. The real failure sits downstream: nobody, anywhere in the chain from record creation to database entry, asks one question — name a football club in this record. That question takes four seconds. It blocks this entire class of error. And it requires no model whatsoever.

What bothers me more: that obituary is, technically, a clean piece of writing. It follows the template order exactly — death, signature roles, awards, other credits, biography, survivors, legacy summary. Assembled at a desk, with no reporter on the ground. Football has the same disease, except we give it prettier names: generational talent, the heir to someone, the jewel of the academy. All of it is newsroom framing, not attributed testimony. In an obituary, "a celebrated screen legacy" is framing. In football, "the best player of his age group" is also framing. Same error, different label.

There is a paradox here: the real value of this record is negative. It is a perfect control case. A good football analysis system must return an empty result when it reads this. If it invents a tactical narrative, the system is broken. If it returns null, the system can be trusted. We usually test systems with positive cases, because positive cases are what we can show off. The negative case is the hard one, because its reward is silence.

If I could change one thing in the pipeline, I would add a post-labelling gate: a record carries the football label only when at least one football entity has been identified. The cost is close to zero. The price is accepting that some throughput gets blocked, and accepting that someone will ask why their report was held back.

My trade is re-reading. Before writing about the future, read today one more time. If you hold a football dataset, pull twenty records at random and ask that four-second question. I did. The result was not pretty. But at least I know which layer of soil I am standing on, and which layer is only aura.