An earthquake report tagged 'football': the hole is at the intake stage

At 9.40pm in the Shenzhen newsroom, I opened a file in the sports desk's intake log. The first line read: domain label — football. The second line read: 16 dead, 32 injured, 339 homes collapsed. Location: Nagqu, northern Tibet. Magnitude 7.0.

I read it three times, slowly, the way I read a match report. There was no club in the file. No player, no coach, no competition, no contract, no league table, no federation. There was an earthquake, a damage assessment, and twenty-three information points numbered one to twenty-three.

A system had tagged a disaster report as football. Across all twenty-three information points, not one could object to that label, because not one of them mentioned football. I sat still for a while. Nine years in the city that manufactures most of the screens the world looks at had taught me that screens say wrong things. This time it said it in a way that made me take out my notebook.

Context: the label is the only door

In 2026, at 46, I left a print newsroom for a digital sports platform. My first assignment was Shenzhen FC against Wuhan Zall in China League One, with 4,213 people in the ground. I still kept a paper notebook; the desk required a live post every three minutes. I objected that I had no verified data, then complied anyway. Thirty rounds later I understood that the gap between fast news and correct news lives inside that three-minute window.

That process is now almost fully automated. A mid-sized sports desk ingests thousands of items a day from hundreds of sources: football, esports, basketball, tennis. Nobody reads them all. A classification model assigns a domain label, a ranking model decides which items deserve a human eye. The label is the only door into human attention.

When that door is mislabelled, two things happen. Either the item lands on the desk of someone without the expertise to handle it, or — far more often — it drops into a folder nobody opens. Both outcomes end the same way: the desk's picture of the world is built on labels nobody re-checks.

The transfer window makes it worse. Dozens of items a day about the same player. The system rewards volume, because volume produces page views, and page views are the only thing measurable instantly. Readers are drowning. What they need is a reliability filter, not more items.

The paradox: the things that actually decide a deal — release-clause structure, payment schedules, the wage bill once bonuses are loaded, performance-contingent terms — are the items with the fewest sources, and therefore the lowest ranking in a volume-driven system. A rumour with five outlets floats to the top. The actual contract sinks. The desk does not choose that; the intake architecture chooses it.

What actually happened: a vocabulary test

Set the label aside and read the entity list: Tibet Autonomous Region, Nagqu, Lhasa, United States Geological Survey, China Earthquake Networks Center, Xinhua, CCTV, local police and fire services. Not one football entity.

A label is a hypothesis; entities are the evidence. This is the test I run on every item before it enters my notebook: list the entities, count how many belong to the field the label claims. On this file, the count was zero. No further reading was required to know the label was wrong.

The troubling part comes after. The system has no step that asks whether its own label is plausible. It assigns and forwards. And because nobody opens the file, the question never gets asked.

Five repetitions, one source lineage

The figures 16 dead, 32 injured, 339 homes appear five times across the twenty-three points, at positions 2, 6, 8, 18 and 23. Skimmed, five repetitions look like five independent confirmations. Traced to origin, all five lead to the same place: official Chinese media quoting Nagqu municipal officials.

This is the single-source echo — one statement republished five times, and by the fifth time it has the shape of a consensus.

Football has a fully developed version of this, and it earns more money. An agent whispers a line to a reporter, the reporter publishes, four outlets cite the reporter, an aggregator cites the four outlets. By morning the fan reads "multiple sources confirm". In reality there was one mouth. I have read hundreds of those chains across nine consecutive transfer windows, and the only way not to be fooled is to count sources at the point of origin, not at the point of republication.

Two agencies, two numbers, and the ban on averaging

The only genuinely independent source in the file is the seismological data. The United States Geological Survey records origin time 22:33 local, epicentre 191 km east-north-east of Nagqu, focal depth 18 km. The China Earthquake Networks Center records a depth of 10 km.

Two agencies, two different numbers for the same parameter. The correct handling, per the process I have used for nine years: record both, flag as pending verification, do not average, do not pick whichever looks more authoritative. Focal-depth divergence is a routine measurement artefact, and two agencies disagreeing is not a contradiction to be smoothed away — it is information.

Football does the opposite almost reflexively. Two data providers give different expected-goals values for the same shot; the desk averages them and prints a figure to two decimal places. Two tracking systems give different sprint distances; the desk picks the prettier one. We call this data professionalism. I call it manufacturing false precision out of two sources that were never cross-checked.

Writing 'insufficient information' is a skill, and nobody pays for it

The correct conclusion of this analysis, across all nine dimensions, is "insufficient information". There is no formation to assess, no player to track, no contract to unpick, no table to compare. An empty analysis is the accurate output, not a failure.

In this trade, empty is rarely praised. People are praised for producing. That pressure creates a reflex: fill gaps with inference, smooth ambiguity into prose. I did exactly that at 46 and caught myself doing it.

The 2026 World Cup is where data betrayed me most clearly. Belgium against Brazil: I analysed Brazil's 78 per cent possession and predicted a Brazil win. Belgium won 2-1 on the counter. I sat down, collected the physical data from twelve knockout matches, and found that possession did not correlate with goal share in that phase of the tournament. Since then I have dropped the habit of quoting one metric and drawing a conclusion.

The 2026 World Cup taught me that data can forecast the future but cannot forecast the heart.

In 2026, inside Guangzhou Evergrande's closed training bubble during the pandemic, I sat in a 58,000-seat stadium with nobody in it and counted Zhang Linpeng take 47 free kicks in 38-degree heat. No sensor recorded the loneliness. No statistical table has a column for the silence between two strikes. Data is a map; the match is territory nobody has ever surveyed.

Since then I keep one rule in the notebook: one match, one scene. After every game I must record exactly one moment no table can reach — a glance, a shout, a back line rising as one. The rule does not make the writing better. It stops me confusing the map with the territory.

Esports is where intake errors do the most damage and draw the least notice. A professional career is shorter than a footballer's, the talent system is thin, post-retirement support is close to non-existent. A 19-year-old who retires with a wrist injury may get exactly one article written about him in his entire career. If that article is mislabelled and drops into an unopened folder, the story vanishes from that sport's history. The archive was already thin, and we are thinning it further with labels.

The contrarian angle: the mislabelled earthquake is the luckiest failure in the batch

The instinctive reaction to this incident is to call it dangerous.

It is not dangerous. It is loud. A report about 16 deaths sitting in a football folder will always be opened by somebody, because its content is too large to ignore. It failed in a way that cannot be hidden.

The frightening errors are silent. A report on a club's wage structure tagged "entertainment" will never be opened. A figure repeated five times looks exactly like a consensus. A metric averaged from two divergent sources looks exactly like a precise metric. All of them pass through the door without a sound, and the desk's picture of the world tilts slightly, and then slightly more, every day.

The industry's reflex is to demand more data, more models, more intake layers. I think that is poison. The more sources read from one mouth, the more their errors correlate instead of cancelling out. Football already walked this road at the tactical level: every club hunts the same inverted-winger profile, every front line converges on the same shape, and once the whole league looks alike nobody gains an edge from being different. Sports data is on the same road. Source diversity is the only error-correction mechanism we have, and we are deleting it for speed.

In Shenzhen I learned that a screen cannot replace the stand. A screen only shows you what the data pipeline has decided you are worth showing.

Signals to track

Three things, and I will do all three this transfer window.

Measure the mislabel rate instead of the model's accuracy. Pull a random sample of ingested items, open them, check the label against the entities. If the rate crosses a threshold the desk sets for itself, stop and fix the intake.

Deduplicate by source lineage. Five information points leading to one agency count as one, not five. Apply the same rule to every transfer chain before publication.

Record both figures when two sources diverge, and say clearly that verification is pending. No averaging. No picking a side.

I keep time for seasons that are already gone, even when nobody is listening any more. That beat now includes checking the labels nobody checks. I never run faster than the match; I only keep time to the last minute. And I will still open every file, read every entity, before believing the first line — because a sports desk is only as trustworthy as the weakest stage of its intake.