An obituary inside a football database: entity mislabelling and the cost of a dirty corpus

On 22 October 2026, inside a corpus labelled football, I stopped at one record. Seventeen information points. No club. No player. No match, no table, no line of a contract. The record told the story of an American actor named Michael Gray, who died on 16 October 2026 at the age of 75, known for playing Billy Batson in the television series Shazam!, which ran for three seasons from 2026 to 2026. The cause was given as natural causes, relayed through his representative and then through US media. That was all. Two decades of reading football files taught me that what is hidden usually sits in the appendix. This time the error sat in the label itself. An obituary was filed in football's drawer, confirmed as football by the system, and would be read back as football by anyone who searched for it later. Nobody had to lie. A string of characters only had to collide. The pipeline has three layers. The first reads text and extracts information points. The second assigns a domain label. The third stores the record for search, statistics and model training. The life-or-death decision sits in the second layer, and it is made on entity names. “Michael Gray” is a string. In football the string exists: an English left-back born in 2026, who played for Sunderland and Blackburn Rovers. In entertainment it also exists: an actor born in 2026, dead in 2026. Four disambiguation keys were available and unused: birth year, nationality, occupation, club affiliation. Two of the four matching would have blocked a false merge. The system blocked nothing. There is a secondary signal I logged at low confidence: the keyword “Captain Marvel”, attached to a comic-book property, collides in some taggers with sports vocabulary. I disclose it for transparency and do not use it as evidence. The evidence is elsewhere: a record containing no football entity at all still carries a football label. Scale does the rest. At a few thousand records, one bad row is a nuisance an editor will catch. At millions of records, used to benchmark models, a mislabelled row with a correct-looking label passes every content check, because content checks assume the label is right. From the 2026 press room to the Girona bot network of 2026, power only changes shirts. The mechanism is unchanged: a system declares something true, and nobody re-checks the layer beneath. FOUR KEYS, ONE TEST In 2026 I took on Real Betis transfer files and found the doping trail not in the summary but in the contract appendix. Betis hid doping in the appendix; I read backwards, page by page, until I found it. That discipline applies to any information system: do not trust the summary, read the fields behind the label. Here the fields behind the label are uniformly empty. A record labelled football with no club field, no competition field, no season field, no performance metric. The simultaneous absence of all four is a stronger signal than any sentence in the piece. When a file is missing every mandatory field at once, the correct conclusion is that the file was misfiled, not that it was thinly written. The test I propose is cheap: hard-match on occupation, birth year and nationality before allowing any merge into a player record. All three keys are present in the source text. An actor born 2026, American, occupation acting. A defender born 2026, English, occupation football. No key matches. The test takes under a second. HOW DIRTY DATA SPREADS In 2026, when Girona had just been promoted to La Liga and sold a 22-year-old defender for ten times his market valuation, I downloaded all 40,000 interactions on the player's account and found 12,000 accounts sharing a single API key. Girona inflated the price with a bot network; the real value sat in the server log. The lesson was not the one the press usually tells: what was inflated was not the price but the credibility of a claim. The mechanism here is identical. An off-domain record labelled football damages three things. First, corpus precision: if two per cent of football-labelled records belong elsewhere, every aggregate drawn from the corpus drifts in an undetermined direction. Second, retrieval: a reader searching for the defender Michael Gray receives an obituary paragraph, and an automated answer system will attach one man's death to another man's living profile. Third, evaluation: if a test set contains mislabelled items, the model is scored against a dirty ruler and the error never surfaces. What troubles me most is the invisibility. A wrong tactical claim can be refuted with video at minute 67. A wrong label has no minute to compare against, because it is generated before the content exists. THE FREE TEST Before checking sources, I check the calendar. 16 October 2026 fell on a Friday, consistent with “died on Friday”. 22 October 2026 is six days later, consistent with the stated disclosure gap. The text's internal arithmetic holds. The test is cheap and I use it constantly, because it filters text assembled from randomly generated numbers. It proves internal coherence, not truth. Those are different categories, and conflating them is the most common failure in this trade. External corroboration I could not do: no family statement, no named confirming body, and the event sits at a point where I hold no public record to compare against. I state that limit rather than filling it with inference. THE TWO-STEP ATTRIBUTION CHAIN The cause of death travelled three hops: representative, then US media, then the article. The originating outlet is unnamed. The image is credited only to a user-editable database, with no licence terms. The survivor details — a 2026 marriage, stepchildren, a son-in-law — are unsourced. Three confidence tiers must be separated. Death: high. The “natural causes” mechanism: medium, as a second-hand single attribution. Family details: low to medium. The correct handling is to hold them as “reported”, never to restate them as confirmed, and to re-verify before any republication. I do not trust transfer fees; I trust the numbers that were crossed out. Here the crossed-out numbers are the empty fields: no named outlet, no named spokesperson, no image licence. Those three gaps say more than the text that was written. BORROWING HEAT FROM A RUNNING FRANCHISE The article anchors a 1970s career to a 2026 film in which the character appears briefly. The modern franchise belongs to another actor's iteration. This is standard obituary technique: hook a faded past to a live asset to manufacture relevance. It works, and it systematically overstates contemporary relevance. Football runs the same mechanism. From my own match-watching experience in England between 2026 and 2026, I remember the Wembley play-off final between Sunderland and Charlton Athletic, 4-4 and decided on penalties, where Sunderland's left-back Michael Gray missed the last kick. A player like that can be recalled for twenty years by a single frame, and the frame is often used in place of the whole career. The same logic appears at the accounting layer. Signing fees for free agents are more toxic than transfer fees because they sit outside the core scrutiny of FFP: the same money, a different label. At the level of the laws, five substitutions deepen a squad but turn the last twenty minutes into a war of attrition; the stat sheet still records one match while the operational reality has changed. The label decides what gets audited, and what drifts past. METHOD NOTE My inputs were the raw text record, a calendar check, and the mandatory-field schema of the corpus. I could not access the original file of the US outlet that reported first. I could not verify the family details. I checked the name-collision hypothesis separately and rate it high confidence, because four independent keys point to two different entities. All conclusions are bounded by the data stated. THE REASONABLE PART OF THE OTHER SIDE One serious counter-argument deserves stating, because it is partly right. Tagging systems exist to scale; if every record needed a human reader, the corpus would not exist. A tagger operating at 98 per cent precision is a technical achievement, and one rare failure does not erase it. The obituary itself is carefully built: a self-consistent timeline, a properly hedged cause of death, no invented detail added for shock. I have also observed exactly one case, and I do not know the base rate. That is precisely why the danger is larger, not smaller. A mislabelled record cannot be caught by content checks, because every content check begins from the assumption that the label is correct. The fault sits on the layer nobody thinks to inspect. I considered filling six empty dimensions with plausible tactical prose, because it would have produced a far more readable article. I did not. A football analysis built on an obituary is a worse failure than a table left blank. One last point, stated plainly: neither man named Michael Gray did anything wrong. The fault lies in the pipeline design, and design can be fixed. Three things to do now, in order. Quarantine this record from all football aggregates and log it as a negative control in a regression suite. Add a sport-versus-entertainment gate upstream of the domain tagger, not downstream. Audit every Michael Gray record already in the corpus for prior contamination. A corpus is not judged by how many records it holds, but by how many it dares to remove. The remaining question belongs to the people running these systems: who inspects the layer that inspects everything else?