The Scoreline Held; the Name Did Not: A Verification Log of Arsenal's Carabao Cup Exit

The screenshot landed on my phone at two in the morning in Barishal, and its headline had already decided everything: “Arsenal’s resounding collapse.” Beneath it, in smaller type: out of the Carabao Cup, beaten 5-3 on penalties. I did not start with the scoreline. When I load any match report into my verification sheet, the first column I fill is not the result — it is the name list: scorers, penalty takers, cards, minutes. Scores publicise themselves. Names are where reports quietly break. This one broke at Bruno Guimarães. The report states that the Fleetwood Town goalkeeper saved his kick. In an Arsenal–Fleetwood shootout, that kick has to belong to Arsenal. But Bruno Guimarães is a Newcastle United midfielder with no team-level connection to this fixture. Two rows from two different tables have been joined, and no error message fired. The scoreline was clean. The list was not. So the screenshot did not read to me as a match report. It read as a quote aggregation wearing the shape of one, with a foreign row smuggled into the dataset. What follows is the log of me chasing that row. The article’s architecture is easy to reconstruct. Its spine is Mikel Arteta’s post-match reaction. The quotes come from an interview with BBC Radio 5 Live, later carried by the outlet Tribuna. Arteta calls it a “very sad day,” praises Fleetwood, declines to question his players’ attitude or commitment, laments missed chances, and accepts the exit. The match summary adds: Fleetwood scored in the 21st minute, Arsenal equalised in the second half, chased a winner, and lost the shootout 5-3, with the hosts converting every kick. The headline says collapse. The body describes a one-step margin in a shootout. That gap is not decorative; it is structural. Call this genre the quote-led match report. It is not a moral failure so much as an economy. A post-match press conference runs eight to ten minutes; from it you can extract six hundred words without requesting a single additional document from the club, and without a reporter leaving the house. The supply chain gets short: radio interview, then wire, then outlet, then a forwarded message. Each hop sheds metadata — date, venue, round, lineups, card minutes, penalty order. A report of a cup upset needs a minimum set of rows: the date; the venue and competition round; both lineups; scorers with minutes; cards with minutes; the penalty sequence with taker names; and both managers’ comments. This article contains none of the first six. On the seventh, it has one manager. I am not writing this to catch an outlet. I am writing it because I have spent years logging this exact failure mode in a different dataset. In 2026, aged twenty-three, working as a junior data journalist at FootballLab BD, I charted Bangladesh versus Afghanistan in the AFC Asian Cup qualifier. Fourteen shots, Bangladesh 0.87 xG, Afghanistan 1.12 — and Bangladesh scored from a 0.08 xG shot. That night taught me that xG is a range, not a verdict. I spent three weeks rewriting the code and adding uncertainty. But that night taught me something else I underrated at the time: the 0.08 goal was real. The roster was real. The uncertainty lived in the model, not in the name column. A report with the wrong names is a different class of failure — uncertainty cannot rescue it, because the problem is not the model, it is data entry. Start with entity resolution. A shootout table typically carries keys: match ID, taker ID, team ID, outcome, kick number. The article’s own arithmetic opens a hole in it. If Fleetwood converted all five and the final score was 5-3, then — assuming both sides took five kicks — Arsenal had to fail twice. But the report narrates exactly one failed kick, and even that one belongs to a player from a third club. So the table has a foreign key and an empty cell. Three hypotheses must stay open. One, the name is wrong. An Arsenal player’s kick was saved, and the name was pulled from an adjacent article during aggregation. Two, the club association is wrong. Two different matches have been spliced; a Newcastle-related shootout line has wandered in. Three, the shootout arithmetic is wrong. Maybe it was not 5-3. Maybe it went to sudden death. Maybe there were more than two failures. Each has a test: the official EFL or club match report, the penalty order, the card minutes. Until one test passes, all three stay alive. Picking the most plausible and calling it verified is the habit that reduces verification to decoration. Name errors are cheap to make and expensive to inherit. A wrong xG gets argued about; a wrong name gets copied. In Bangladesh Premier League coverage I have watched a single roster error travel through five outlets in a day. A tactical opinion has a half-life of days. A name error has a half-life of months. The second row is heavier, and it sits inside the article itself — just not inside the match summary. Arteta is quoted saying Arsenal played with ten men and then conceded a poor early goal. The summary contains no red card, no player name, no minute. When a material fact appears in the framing but not in the record, one of two things is true: the summary is incomplete, or the quote was mangled somewhere in the chain. Why this row carries the most weight is obvious in game-state football. A red card in the 20th minute means seventy minutes of ten against eleven. Everything downstream — possession, shot counts, PPDA, xG — has to be read conditional on that. Without it, comparing xG is dividing by the wrong denominator. In the 2026 World Cup semi-final I ran a live xG model on Croatia–England: after 120 minutes England 1.82, Croatia 1.54, Croatia’s PPDA 8.9. The number alone said little; what it described was a midfield pressing pattern. At Euro 2026, Italy survived Spain’s 1.53 xG with 0.73 of their own, PPDA 13.8 against 6.2 — again a game-state story, not a finishing story. At the 2026 World Cup, Japan beat Germany with 26 percent possession and two shots on target against 1.87 xG. All three examples share one condition: some game-state variable was logged. Here the red card exists in a quote and nowhere in the dataset. My rule is simple — a claim that would change the entire reading of the game state must carry its own evidence row. Otherwise it is not a quote, it is a rumour with a voice. Third, the arithmetic of the headline. Collapse implies a set of things: a wide margin, systemic breakdown in the face of the opponent, perhaps one-sided shot numbers. The body offers a 21st-minute concession, a second-half equaliser, a chase for a winner, and a shootout defeat. That is a cup shock, not a collapse. Run a back-of-the-envelope, with every assumption labelled. Suppose both sides convert at roughly 0.75 per kick. I am not offering that as a measured conversion rate for these two teams — the report provides no such data — only as an anchor to show the shape of the problem. Under that assumption, a five-kick shootout settles around 4-4 or 5-4 most of the time, decided by one kick. A 5-3 is a two-kick gap: not nothing, but it is one or two saves, not a rout. For collapse to hold, we would want something like 4-0 with 0.4 xG against 2.8. Nothing of that sort appears here. My objection is not to emotion. Emotion is an editorial choice and a legitimate one. My objection is to measurability: a headline should make a claim the body can falsify or support. Collapse is not falsified by the body — only softened. A claim outside the reach of verification is not journalism, it is mood. Fourth, the provenance chain. The quote path runs: BBC Radio 5 Live, then Tribuna, then this article, then a forwarded message. No transcript, no audio link, no timestamp, no byline. Match facts are sourced to the article itself — a self-referential citation. There is another signal that matters as much for data integrity: an unrelated Messi and Ballon d’Or link is stitched into the middle of the match report. That kind of insert is usually placed automatically by entity proximity. When the link has no relationship to the fixture, one thing is settled — nobody who read the article assembled the page. And the largest absence: there is no Fleetwood voice. Arteta gets six or seven quotes; Fleetwood’s manager and players get none. That is not neutrality; it is one microphone described as the whole room. A clean dataset can still lie when the crowd is missing. Still, credit where it is due, and it should be said before the criticism hardens: the article’s account of Arteta’s reaction is internally coherent. The sad-day framing, the praise for the opponent, the refusal to blame his players, the frustration over missed chances, the acceptance of elimination — that is plausible press-conference behaviour. The failure is not in the emotional inference. It is in the rows. Which brings me to my most uncomfortable observation. A wholly false report is easy to catch. A partially true one is dangerous. The story that Arsenal lost a shootout and Arteta was unhappy is coherent, natural, and probably accurate. That coherence is the transmission vector. Readers do not stop after the first paragraph; they enter the structure, and the smoother the structure, the fewer questions they ask. A wrong name and an unexplained red card stop registering as anomalies. They become part of the story. I should also name my own bias here. When a number smells wrong to me, my instinct is to discard the model and rebuild it. But a rebuilt model is not a correct one. The rebuild log and the validation log have to be kept separately; a new model is a hypothesis, not a verdict, until it survives out-of-sample matches. The same applies here: I am not claiming the match did not happen. I am claiming this account cannot be cited verbatim. One more thing — headline exaggeration is not a moral failure, it is an optimisation. What travels is the emotional label; what gets checked is the scoreline. A page that does not distinguish between the two is not deceiving its readers. It is feeding them exactly what they asked for. The problem is that what we want is not always what we need. Save the sharpest observation for last. The sentence about playing with ten men is, if true, the most analytically valuable sentence in the entire piece. That is precisely what makes it costly. Trust always migrates from the most checkable claim to the least checkable one. A reader sees a date line up, feels satisfied, and then believes the red card too. Looking ahead, I have three signals that will tell me whether this page got repaired. First, whether official channels — Arsenal, Fleetwood, or the EFL — publish a report containing card detail. That resolves the ten-men row and tells us whether the summary was incomplete or the quote was corrupted. Second, whether the Guimarães line is corrected with a note or silently deleted. A correction with a note is a repair. A silent deletion is an admission without a timestamp. Third, whether a byline, date, and venue appear. Provenance either gets attached or the page keeps floating — and floating pages are the ones that get forwarded fastest. Every quote is a variable waiting for a timestamp. Tonight I wrote down only this: the scoreline was clean; the list was not. The question is how many other rows in my feed carry the same fingerprint — and how I will find out, if not after someone repeats them first.