Zero Information Points, a Football Label: How a Broken Analysis Pipeline Makes Newsrooms Fail Silently

On a Monday morning a file landed on the desk with a green tag in the corner: football. Everything beneath it was empty. No title, no source, article type marked unclassified, the information-point array blank, the core-viewpoint field white, no entity identified, time sensitivity unassessed. A document routed for nine-dimension deep analysis arrived with zero analysable substance.

I spent nine months inside Tottenham during the final season at White Hart Lane: 38 training sessions, 19 away trips, a recorder in the mixed zone, fan-group WhatsApp threads until one in the morning. Harry Kane scored 29 league goals that year; the side finished second on 86 points and went unbeaten at home. But what took up more room in my notebook than those numbers was a small habit. At the end of each session I used two different pens. One wrote, nobody said anything today. The other wrote, the notes are lost. Both produce a blank page. One is information; the other is the absence of it. The old ground taught me that a goodbye has a tempo you can only hear live, and before you can measure a tempo you must decide whose silence you are measuring.

The file that arrived on Monday was the second kind.

Set the scene. On 11 July 2026 in Moscow, England lost 2-1 to Croatia after extra time; Harry Kane took the Golden Boot with six goals. Across 32 days in Volgograd, Nizhny Novgorod and Moscow, every match notebook held at least a name, a scoreline, a kick-off time. That is a live information point. The document on the desk holds none of it. It holds a label that says the subject is football without saying which football.

An empty information-point array and a genuine no-news finding are different records; mistaking one for the other is the real cost of this failure.

Now the context. Over the past decade, the part of sports-desk work machines have taken over is deconstruction: breaking an article into its information points, viewpoints and entities. The reason is simple. To handle the flood of AI-written sports copy, every claim has to be broken down and checked. That breakdown is Stage One. Stage Two is the nine-dimension pass: tactics, club finance, league position, governance, dressing room, risk, media narrative, industry transmission.

The framework is sound, and every one of its dimensions rests on a single condition: the information points from Stage One. With zero information points, analysis is not possible. It can only be fabricated. That distinction is the whole story here.

The document admits as much itself. Its pre-check section states plainly that every measurable field is empty or unavailable, and that only the domain label survives. The artefact is therefore not football analysis. It is a specimen of a broken football-analysis pipeline.

So where did the content go? The document offers three explanations, ordered by likelihood. First, the source fetch was blocked, by a paywall, bot protection, or a JavaScript-rendered page, returning an empty body. Second, the deconstruction step timed out or failed silently and emitted a default empty schema. Third, the input was never an article at all, but an image, a video embed or a redirect page with no extractable body text. None of these is the failure mode of an article that simply contained too little football. An ordinary match report leaves behind at least one scoreline entity.

The combination of a populated football label and an empty body is the most dangerous part, because automated routers act on the label, not on the body.

The document makes a sharp observation: the survival of the word football means the classifier received something football-related, yet the text deconstruction returned an empty array. Two separate steps, one topping the other. A classifier running on metadata and a deconstructor running on body text; the success of one masks the failure of the other. The casualty is a document that looks benign and reads as nothing was found.

In 2026 I watched fourteen behind-closed-doors matches from inside grounds after Project Restart. Tottenham finished sixth. Canned applause played, and the hush between bursts of it was not canned. The empty stadium revealed that silence is not absence; it is a held note. Writing the six-part Empty Lane series taught me that silence has at least three varieties: someone declined to speak, someone could not speak, and someone was never in the room. Newsrooms make their worst errors when they confuse the second and third.

That is exactly what happened here. Nothing was found and nothing ever arrived look identical in an analytical table.

In an analytical table the phrase not available reads well, looks rigorous, and is trusted precisely for that reason.

The document names the trap outright: silent null-finding contamination. An empty deconstruction can be read downstream as the article contained nothing notable, when in fact nobody read the article. Once that happens, journalism stands on a conclusion with no existence behind it, wearing a seal of procedural rigour.

The recommendations are cold-headed and they are the useful part. One: the record must be tagged pipeline failure rather than no findings, and blocked from any reporting or briefing output. Two: the original URL and a raw payload hash must be logged at ingestion so a re-run is reproducible. Three: routing must be gated on the information-point count exceeding zero, never on the domain label. Four: if more than five per cent of records in a batch return empty information points, it is a systemic regression rather than an isolated incident; and for JavaScript-dependent pages, whether roughly 500 words of body text can actually be extracted has to be checked separately.

Those four conditions are not mere process hygiene. They are the credibility architecture of football coverage. The modern part of football journalism rests on numbers, and numbers can be checked. Expected goals, passes allowed per defensive action, transfer amortisation, sell-on clauses, the final contract year; every one of those terms hangs on a calculation. A pipeline that cannot retrieve the calculation brings back a debt instead. The debt is repaid by other people's names on other people's copy.

The transfer market shows this most clearly. A rumour has three layers: the tier of the source, the motive of the agent, the actual need of the club. If the source has no name in the pipeline, that layer is zero, yet the router still pushes it forward because the label says football. The same holds in tactics. Possession share and counter-pressing are fine, but no framework can show the empty spaces in a passing network unless it holds an information point about them.

The event that does not occur in a newsroom is often the largest event, and inside automated systems what does not occur becomes what has been decided.

Now the honest part. Almost always the discussion opens on a single fear: AI is making things up, hallucination. Here the opposite happened. Nobody invented anything. Nobody pulled anything in either. A document came back empty-handed, and its manner of returning was so polite, so restrained, that walking past it is entirely natural.

But an engineer who runs this pipeline would argue differently: this is mainly a cost question. Fetch the source again, re-run the step, a few minutes, close to free. The argument is not foolish. In most cases a re-run does produce a result.

One objection remains and I will not sweep it away. A re-run can explain a failure and can reduce how often it recurs, but it cannot mend an unknown provenance. A document with no title, no source and no originating nature will return the same way however many times it is fetched. On that point I disagree with the operational view, and I am leaving the disagreement open, because what a process owner should prioritise over a reporter is not mine to decide.

A short addition. Reader trust is built from the traceability of every claim. Which source, which date, which number; know those three and anyone can verify a piece. In the failed record all three are unavailable. The reader cannot verify it, and cannot reuse it either. That leaves two options: borrow it or swallow it.

The fixture list never tells the whole heartbeat, which is why I write features. A fixture list gives only time; a feature gives the intervals, who is tired, who went home, who did not. This record is an interval too: nine dimensions, a structure of some 3,500 words, and an empty room inside.

Over the next six weeks I will watch four signals. One: whether a re-run on the raw source returns at least three information points and one named entity. Two: whether the original URL is readable at all; if more than 500 words of body text extract, the cause was a fetch block, and if not, the input was never an article. Three: what share of records in the same batch come back empty; beyond five per cent it stops being a casualty and becomes an outage. Four: the label-to-body ratio; if the rate of empty documents carrying a domain label keeps rising, the classifier is running on metadata alone.

I have a rule of my own, inherited from those two pens in 2026: never erase a blank page, write on it why the page is blank. A document that unknowingly tells you nothing was there is less honest than one that says nothing arrived, send it again. If this failure stays quiet through the rest of the season, later readers of tables, working notes and briefs will never work out what the waistcoat was for, or where the music started.