The file labelled 'football' and the audit of 3.5 million records

The file labelled 'football' and the audit of 3.5 million records

6:40 a.m.

6:40 a.m. on a Tuesday, a fourth-floor flat in the Prati district of Rome. The newsroom's automated classifier drops a text file into my inbox. Label: football. Confidence score: 0.91. I open it. Four minutes later I have read all 37 information points and found no team of any kind.

No player. No coach. No competition. No contract, no season, no league table, no match. The content is the ENTI household survey carried out by Inegi, Mexico's national statistics institute, publishing 2026 data on child labour: 3.5 million people aged 5 to 17 worked during that year.

I sit for another twenty minutes, not to find a way to turn this file into a football piece. I sit to understand how a socio-economic dataset cleared the filter of a sports desk with a 0.91 confidence score. Once I understood, I realised the story worth writing was exactly there.

A filter does not read, it matches

A content classifier does not understand an article. It counts. It sees the words "minors", sees "Mexico", sees a phrase close to "academy" or "youth development", and it assigns a label. This is pattern matching, not verification. A 0.91 confidence score does not mean the content is on topic; it means the vocabulary pattern sits tightly against some category in the training table.

Pinamonti entered my life through a spelling error. In January 2026 I wrote "Andre" instead of "Andrea" for a defender, and my editor made me rewatch three rounds of match footage over three weeks. That taught me something everyone who works with data knows but few say out loud: the smallest error usually points at the weakest joint in the system. A typo led me to a transfer. A misclassification today leads me to a larger question about how the football industry consumes data.

In the 37 information points of this file, three operational fields were left blank at the very first stage: the entity list still contains the instruction line instead of data; time sensitivity reads "not assessed"; source quality reads "judge from the source fields". Those three empty boxes are not a technical detail. They are the traces of a pipeline running without a human check.

Why this belongs on a sports page

I write about the transfer market. My job is to read a number and ask where it came from. When a newspaper writes "club X has agreed a fee of 20 million euros plus 5 million in add-ons", the first thing I do is lift the number out of the sentence and rebuild its route: who said it, to whom, when, and what does the speaker gain if that number appears in print.

In 2026 I tracked 47 rumours involving Italian players during the Russia World Cup and found that 83% of the sourcing had been inflated by the agents themselves to lift valuations before the summer window. In 2026 I was pricing rumours. Now rumours price me. But the method has not changed: a number only means something when you know the denominator, the comparison point, and who is standing behind it.

The Inegi file is a perfect stress test for that method, because it comes from a high-credibility producer and still contains enough gaps for a careless reader to reach a wrong conclusion. That is precisely why it belongs on a sports page: the football industry consumes this class of data every day, at a speed far beyond its capacity to verify it.

The first subtraction

The file's headline table has two points. 2026: 3.7 million people aged 5 to 17 in work. 2026: 3.5 million. A gap of 0.2 million, roughly a 5.4% fall over three years.

The arithmetic is right. The interpretation is not.

I break it down further. Girls: 1.5 million down to 1.4 million. Boys: 2.2 million down to 2.1 million. Added together: 1.4 plus 2.1 equals 3.5. The internals reconcile. For someone who audits data, that is the first good signal: the table does not contradict itself.

But one line barely moves. Domestic chores under inadequate conditions: about 1.9 million, essentially unchanged between the two measurements. This is where I stop longest, because it is the only item in the whole file that the article itself puts forward as proof the problem is not confined to paid employment.

If 1.9 million of a 3.5 million total is housework, then more than half the volume of child labour in Mexico sits beyond the reach of a labour inspector. There is no factory to inspect. No employer to fine. No timesheet to cross-check. And no indicator in a transfer report that could ever reflect it.

The missing denominator

The article reports a rate of 12.6% for 2026. The article does not report the rate for 2026.

This is the largest gap in the entire file, and it is not the writer's error. It is how statistical news is written: take the most striking figure at the newest point, set it beside the absolute figure at the older point, and call the difference a trend.

Absolute headcount can fall for two entirely different reasons. First, fewer children are working. Second, Mexico has fewer children. The 5-to-17 cohort is shrinking along the demographic transition curve of the whole region. If that cohort fell 1 to 2% between 2026 and 2026, then a 5.4% fall in absolute headcount can coexist with a rate that is flat or even slightly rising.

I try the reverse calculation. 3.5 million divided by 0.126 gives roughly 27.8 million people aged 5 to 17. That is a demographically plausible figure for Mexico. It means the numbers in the file are internally consistent. But internal consistency does not create the missing comparison point.

A headline saying "slight decrease" without the rate at both measurement points is an unverified headline, even when the subtraction is correct.

Three scenarios instead of one conclusion

I always write three scenarios, even when only one looks obvious.

Decline scenario: the 5-to-17 cohort shrinks faster than the number of working minors, meaning the true rate rose. The "slight decrease" message then conceals a step backwards.

Flat scenario: the cohort shrinks at roughly the same pace as the working-minor count, meaning the rate stands still. Three years then produced no structural change, and every claim of progress is a political claim.

Recovery scenario: the rate genuinely fell, but the fall is concentrated in the easily measured paid-work segment, while the domestic-chore segment is unchanged. Improvement on paper arrives alongside stagnation on the ground.

These three scenarios do not exclude each other at national level. They can all be true in three different states. That is why I never write a single conclusion from a national dataset.

Shocks and an absent baseline

One of the most striking data points: 83.3% of households with working minors had experienced at least one serious shock in the previous three years.

That figure is powerful. It is also very easy to misread.

The problem is the absence of a control group. The file does not say what share of Mexican households in general experienced a serious shock over the same period. If the national baseline is 80%, then 83.3% says nothing about causality. If the baseline is 40%, then a 43-point gap is a major finding.

I do not have that baseline. So I do not conclude.

What I can say is that the structure of the shocks is very clear. Food prices: 48.4%. Loss of paid-work income: 13.9%. Drought or lack of rain: 9.4%. Those three figures form a meaningful order: the cost-of-living channel is roughly four times larger than the job-loss channel.

For a market analyst, the order matters more than any individual figure. It says the shock reaches households mainly through prices, not through employment. That implies income support and food-price stabilisation may be more effective than pure job-creation programmes. A policy conclusion drawn from the ordering of three numbers, not from the value of any one of them.

35% receive nothing at all

Another data point breaks the standard reading: 35% of minors in non-permitted occupations receive no payment at all. A further 8% are in own-account work, a form of household micro-enterprise.

If a third of working minors are unpaid, then the model of "child labour as an income strategy for poor households" explains only two thirds of cases. The rest is labour substitution or family obligation: the adults go out to earn, the children stay home carrying work nobody pays for.

This is where I think of Arthur and Pjanic. Arthur-Pjanic taught me that a deal can die on the pitch and still live on the balance sheet. A swap valued at 72 million euros plus 10 million in add-ons, in reality a book-balancing device in a season when revenue fell 45%. On the pitch, two players swapped shirts. In the accounts, two profits were booked at once. Two realities, one transaction.

The Inegi table operates the same way. On the books, 3.5 million people are working. On the ground, more than half of them are doing work nobody records, nobody pays for, and nobody inspects. A clean balance sheet conceals a different reality.

The 49.1% versus 4.6% gap

The most shocking data point in the file: 49.1% of minors in the working group do not attend school, against 4.6% in the non-working group.

A gap of more than ten times. Very easy to turn into a headline. And very easy to use wrongly.

This is a cross-sectional association, not a causal relationship. Causality can plausibly run in both directions, and it can run in a third direction nobody names: poverty, distance to school, instruction not delivered in the mother tongue, missing birth documentation. A child without papers cannot enrol, a child not enrolled stays home, a child at home takes on work. Child labour may be a consequence of not attending school rather than a cause.

The dataset cannot adjudicate that. It only places two numbers side by side.

For anyone writing about markets, this is the most common and least detectable error: reading an association as a cause, then building a complete narrative on an unverified relationship. I did exactly that with Calafiori. In July 2026 I correctly predicted the move to Juventus at 50 million euros plus 5 million in add-ons, publishing three days before the official announcement. But I ignored that Bologna lost three key players at once, and the side took only 9 points from the first 10 rounds of 2026/25. My readers said I could see the tree but not the forest. They were right.

That lesson applies intact here. A correct deal does not mean a correct analysis. A correct correlation does not mean a correct cause.

Guerrero and Oaxaca: named, not quantified

The article names Guerrero and Oaxaca as the two states with the highest rates.

I went back through all 37 information points looking for the rates for those two states. There are none. No rate, no ranking, no sample size, no confidence interval.

This is the type of gap that irritates me most, because it looks like information has been supplied. A name in a headline creates the impression the problem has been located. But nobody can act on a name without a number.

What I can do is cross the sector structure against the regional structure. Agriculture accounts for 33.2%, small-scale manufacturing and services for 14.1%. Drought appears as a shock at 9.4%. Guerrero and Oaxaca sit among Mexico's lowest-income states and among those most exposed to climate risk. Those three pieces fit together, and they point to a structural pattern rather than a cyclical one.

But I have to be explicit: the observation about income levels and climate exposure in those two states comes from outside the file, not from it. If I do not say so, I am doing precisely what I criticise.

The gender split: symmetric in total, asymmetric by sector

The totals are near-symmetric: 1.4 million girls, 2.1 million boys. Girls are fewer, but not by a wide margin.

The sector structure is something else entirely. Construction: 9.7% for boys, 0.1% for girls. That is the most extreme asymmetry in the whole file.

A rate of 0.1% does not describe a labour reality. It describes a barrier. No construction site in Mexico hires female minors at a rate approaching zero by natural means. That figure reflects near-total exclusion from formal sites, while the equivalent labour by girls takes place as support for building and home repair inside the household, where there is no payroll and no statistic.

This is why I distrust any analysis that uses totals only. A national policy built on 1.4 million and 2.1 million will design the wrong instrument for both groups. Boys need inspection of construction and agriculture. Girls need intervention in family supply chains and commercial services, where no inspector ever sets foot.

One further note: the female share of the unpaid category is very likely above the 35% average, because female labour concentrates in commerce and services, where family-labour arrangements are most common. But the dataset has no gender-by-payment cross-tabulation. So this is a hypothesis, not a conclusion.

A term nobody defines

The file uses the phrase "non-permitted occupation" at several points. No point defines its legal threshold.

This is more serious than it looks. "Non-permitted occupation" is a legal-statistical term of art, not a general description. It implies the existence of a threshold separating permitted adolescent work, typically from age 15 with restricted conditions, from prohibited work.

The article reports statistics about an illegal activity without citing the law that makes the activity illegal. The reader cannot determine what "non-permitted" means. And that damages the ability to interpret the 49.1% versus 4.6% comparison, because part of the work done by 15-to-17-year-olds is lawful and part is not.

I found no reference in the file to Mexico's federal labour law, to ILO Conventions 138 and 182, or to the ENTI methodology itself. Those three documents are the precondition for anyone using this table for enforcement purposes.

There is a subtler point. ENTI almost certainly uses a legal reference date to define "non-permitted". If the legal threshold changed between 2026 and 2026, the two measurements are no longer comparable. The file does not confirm this. So the entire 3.7 to 3.5 subtraction rests on an unverified assumption.

Nobody is quoted

I read through a third time and noticed something that appears in no table.

The file quotes no official. No minister. No state governor. No NGO. Not one minor. Not one household.

All 37 information points carry Inegi, ENTI, or the article's author as source. This is an editorial choice with consequences: it depoliticises the story. There is nobody to blame, nobody to rebut, nobody to argue with.

On ethical grounds, not quoting minors is a sound safeguard rather than an oversight. Not quoting any responsible authority is a different matter. It leaves the table adrift: anyone can use it, nobody is accountable for how it is used.

And the framing deserves attention. Households are described through shock exposure (83.3%), not through choice. That framing locates responsibility in economic conditions rather than in the conduct of parents or employers. It is a choice with a direct bearing on which policies get proposed: income support rather than enforcement.

The contrarian angle

I sat with this file and asked myself: where is the actual failure?

The easy answer is that a socio-economic dataset entered a football content pipeline. That is an operational failure; fix one line of classifier code and it is done.

But saying only that skips the harder part. The mechanism that put this file in the wrong place is the same mechanism that puts most numbers in the transfer market in the wrong place. A pattern-matching filter, a writer who does not check the denominator, an editor with no time to cross-check, and a headline optimised for speed rather than accuracy.

In the transfer market this failure mode has its own vocabulary. A fee quoted without the payment structure. A wage quoted without the contract length. An add-on dropped from the sentence because it makes the headline figure less impressive. Every time, the reader receives a fact that is arithmetically true and semantically wrong.

The cheapest rumour is the rumour we most want to hear. And here, what we want to hear is that things are getting better. 3.7 down to 3.5. Slight decrease. Slow progress, but progress.

Three years, 5.4%, against a base of 3.5 million people. That is not a structural break. It is a flat line with noise. And if the underlying rate is flat, then the "slight decrease" message has done something more dangerous than being wrong: it has reduced political pressure at precisely the moment political pressure is the only thing that can change the structure.

I no longer chase breaking news. I chase the reason breaking news gets lit. The reason here is obvious: a statistical release needs a headline with a direction. Downward is the easiest direction to sell.

What falls next

I closed the file and wrote three lines in my notebook.

Line one: ask operations to block the football label for any content containing "household survey" or "national statistics institute". A filter that only matches patterns will always fail; the operator's job is to place a gate where the failure does the most damage.

Line two: if this file returns to the pipeline as source material for analysis, it must travel with three documents — the ENTI methodology, the legal threshold for "non-permitted occupation" at the time of the survey, and the 2026 child labour rate. Missing any one of the three, the analysis should stop at description.

Line three: a note to myself. Next time a clean dataset arrives under a headline with a direction, the first task is to find the missing denominator. Not to reject it. To know what I am reading.

3.5 million people are still there, most of them doing work with no payroll, no inspector, and no name on any team sheet. The next question is not how far that number will fall by 2028. The next question is who will be first to publish the rate at both measurement points, so that the subtraction finally means something.