When a Pet-Registration Record Walked Into a Football Dataset

On a recent morning I opened a Stage-1 file. The domain label said, plainly, football. I was working through wage structures and pressing fit inside a transfer window, which is the ordinary shape of my day. What the file contained had no football in it at all: a public-service explainer about the registration of dogs and cats in Mexico. Twenty-two information points. Every one of them concerned companion-animal registries, state legislation, and a bill before the Senate. No club. No player. No match. No transfer fee. I read the gap between label and content twice, because my first reflex was to assume I had opened the wrong document. The second read made it clear the error was not mine. The spreadsheet never lies, but it often whispers. That morning it was whispering in the wrong room. It is worth stating how a modern sports-data pipeline works. Each document is machine-read, then assigned to a vertical: football, cricket, basketball, policy, health. That assignment is the domain label. A correct label lets the document enter its own house; a wrong one leaves it standing at someone else's door, and nobody notices. Half my job is standing at that door, checking who belongs inside. The document itself deserves a fair reading, because its content is genuinely useful. A rumour has been circulating in Mexico that pets now require a CURP. CURP is Mexico's unique population identification code. The explainer clarifies that no federal mandate exists yet. Some jurisdictions have their own systems: Mexico City operates a registry called RUAC, and the procedure there carries no cost. Nuevo León provides for registration under its animal protection and welfare law. A national companion-animal registry bill has been proposed in the Senate, its framework still being defined, while states continue under their own rules. The author's stance is neutral; the purpose is to inform, and to spare people avoidable fines. The problem is not the document's quality. It is the architecture around it. Contamination travels through three layers. First, entity extraction. The model pulls tokens from Spanish administrative prose: registro, transferencia, ventana. Spanish football journalism uses the same vocabulary — transfer, window, registration. When the tokens match, the model mints a false node with no real player behind it. Second, sentiment aggregation. A neutral explainer's tone folded into a football sentiment index starts carrying a false signal. Third, retraining. False positives accumulate, the model learns from its own error, and the error settles in permanently. Years of watching matches taught me a habit: I check the pressing line before the scoreline. That habit is also why I know a dirty input never yields a clean output. In 2026, writing the Expected Value newsletter, I flagged Roma's Mohamed Salah as undervalued — 15 Serie A goals, 11 assists, 2.8 shots per 90, 13.9 xG, 8.7 xA. Every one of those numbers depends on a row sitting under a specific entity ID. If the row sits under the wrong node, the analysis looks elegant on paper and means nothing in practice. In 2026, with stadiums empty and budgets collapsing, I recommended Diogo Jota from Wolves at £41m — seven league goals, 6.1 xG, 2.1 shots per 90, 7.9 PPDA. My Crisis Transfer Index combined wages, age, injury history and pressing fit; it was only as reliable as the rows feeding it. When the stadiums emptied, the models had to learn to breathe — but before a model breathes, it needs to know which room it is standing in. In 2026 in Qatar, Sofyan Amrabat's numbers — 4.1 tackles plus interceptions per 90, 90 percent pass completion, 7.2 progressive passes — were cited by two European recruitment departments. Had the foundation been contaminated, that citation would have been a liability rather than an asset. This is where CURP becomes relevant, and it is my actual finding. CURP is a unique-key system: one person, one code, no duplication. A registry demands uniqueness. A classifier is the opposite — it produces probability distributions, and ambiguity sits inside its tolerance band. Football analytics has spent a decade building models to price players and almost none to price record identity. We need a CURP for players: a canonical entity ID that places a record once, in one place, under one name. The risk type is clear. It is systemic, an information-integrity risk — high likelihood, high impact. A bad record does not merely stay wrong; it casts shadow on its neighbours. Suppose one such error enters a store; if it links to a wage row and sits beside an injury record, it returns multiplied in the next index report. The minimum response: quarantine the record, correct the label, audit the classifier, and sample-check recent records tagged football. The reflex is to fix the label and move on. Before blaming the classifier, three alternatives need testing. One, incidental token overlap: Spanish administrative language and Spanish football language share a vocabulary. Two, a feed-routing error: a mixed feed can drop a document into the wrong branch. Three, a human tagging mistake. I cannot yet prove which occurred. This is where I stop myself, because building a rate without a number is the biggest trap in my trade. I do not know this pipeline's false-positive rate — one in a thousand, or one in five hundred. Guessing would not be analysis; it would be decoration. A second point deserves saying: the document is not bad journalism. Where Mexican citizens face fines built on misinformation, a neutral explainer does real work. Its only fault is that it is not one of ours. The more uncomfortable observation is this: we have built models to detect player weakness and none to detect weakness in our own intake. There is a system for finding defensive midfielders. There is no system for finding bad data. Russia taught me that noise travels farther than signal. A file about pet registration in Mexico landing in our football store is another instance of that lesson. The question now is this: how many unfamiliar guests are sitting quietly in our datasets, still counted among the players?