The Empty Template: A Post-Mortem of the Sports Desk's Automated Analysis Pipeline

There were twelve fields. All twelve were empty.

Article Title: N/A. Article Source: N/A. Article Type: Unclassified. One-sentence Summary: blank. Author Stance: N/A. Article Purpose: N/A. Information Points: empty list. Entities Involved: empty. Source Quality: undeterminable. And beneath all of it, the only line carrying a value — Domain Label: football.

It was Thursday afternoon. Rain kept stopping and starting over Liverpool, and I was sitting at my desk with a document that looked, on the surface, entirely full, and read, on inspection, entirely empty. The schema had done its job. The content had not. I had been handed the form of a document, not the document.

I have been writing match reports for fifteen years and this is not the first file of its kind I have held. A club once sent me a "detailed financial report" that consisted of a header, a footer and thirteen blank columns. A tournament organiser once sent a "full draw" with the dates filled in and the teams missing. Both times I had the same feeling I had this week: somebody completed the process and assumed the work had been completed with it.

This is not a report about football. It is a report about the apparatus that claims to report football, and which on Thursday returned a file that was perfectly valid, perfectly error-free and perfectly hollow.

Context: what Stage-1 is, and what it promises

In any modern sports desk, copy passes through five or six layers. The first is the least discussed and the most trusted. It is called Stage-1: extraction. Its only job is to lift a piece of raw source text and flatten it into structured fields — title, source, type, one-sentence summary, author stance, information points, entities involved.

Stage-2 sits behind it and writes the analysis. The entire arrangement rests on one assumption: that Stage-1's output is full. When it arrives empty, Stage-2 has nothing to sit on.

I have a clean comparison point. In 2026, during the Russia World Cup — the first VAR World Cup — I logged all sixty-four matches from a shared house in Liverpool: every review, every overturned on-field decision, the record twenty-nine penalties awarded, and the first VAR penalty in a World Cup final, Perišić's handball, France 4-2 Croatia, Luzhniki, 15 July. I wrote one rule into that spreadsheet on day one: an empty cell means "not yet checked," never "zero." Read empty cells as zeroes and that tournament's penalty count comes out at twelve instead of twenty-nine.

The rule's root is in 2026. Kolkata, Salt Lake, the U-17 World Cup final — England 5-2 Spain before 66,684. Across fifty-two matches I logged every card and every handball call into a notebook, because I wanted to understand why referees from three confederations were enforcing the same Law 12 in three visibly different ways. Some lines in that notebook say, plainly, "did not understand this call." That is an incomplete record, not a wrong one.

The distinction matters. An incomplete record tells you what is unknown. An empty template tells you everything is known and there simply is nothing there.

The audit: four places this could have broken

The first instinct on seeing a blank Stage-1 output is that the model was bad. In my experience, eight times out of ten the model was never reached. The check order writes itself.

One: extraction failure. The scraper captured a rendered app shell, or the content container selector missed, and the text that entered the tube was either empty or made of navigation menus and cookie banners. This happens daily and usually goes undetected, because the system does not crash — it goes quiet.

Two: truncation. A length calculation gone wrong, or an offset out of bounds, or a token budget that mis-billed itself as zero. I had to learn to recognise this shape during the 2026-21 ISL, played entirely behind closed doors across three Goa venues while I covered it from Liverpool — roughly 140 match reports off broadcast feeds at 2 p.m. UK kickoffs. When a feed arrived ten minutes late, I had to log it. Truncation is the same thing: the record keeps its opening, loses its ending, and files no complaint.

Three: the fallback branch. A model call times out, rate-limits or returns non-JSON, the system drops to the default skeleton, and because the skeleton is valid, no error is raised. This is the dangerous one. Here the system looks successful every single time.

Four: field mapping. The content arrived, was full, and the keys mismatched on serialisation. What reads as "unavailable" is actually "untranslated."

Telling these apart is not a curiosity, because the fixes differ. The first two are cured at the ingest gate. The third and fourth are cured in the specification, not the prompt.

The single most useful diagnostic is character count at the point of entry. Near zero, and it is one or two. A return that parses to empty, and it is three or four. I have never once regretted logging a pipeline from day one, and every time I skipped it I started again from zero four days later.

Now look at what this file actually does well. It got nothing wrong. The envelope is intact, the structure is correct, the schema is valid, the status code is 200. Downstream consumed it without complaint. A system that records its own failure as a success is not a system; it is a habit.

Indian football's administrative record is full of that habit, which is probably why the file felt so familiar. In August 2026 FIFA suspended the AIFF. That suspension was not foreign cruelty — the Supreme Court ordered the election the federation had stalled, and a normalisation committee followed. In the language of documents, the fight was never about a person; it was about a clause.

And where documents are missing, numbers arrange themselves. The 131,781 recorded at Salt Lake in 2026 remains the last figure Indian football can actually verify. Most of what circulates after it comes from press releases rather than turnstiles. Sunil Chhetri's 95 international goals have spent twenty years masking a period in which no striker pipeline was built. The closed franchise years from 2026 to 2026, visibility bought at the top and rot left underneath — that line is not an opinion, it is what the ledger shows.

The contrarian angle: the model is innocent; the schema is guilty

The easy reaction is that the AI lied. That file never claimed to be intelligent. Its builders asked the model to return a JSON object of a given shape. The model did exactly that. The shape is correct. The content is zero.

The failure sits elsewhere. Nowhere in the spec did anyone write down what should happen on "empty." Nobody defined how empty counts as empty. Nobody said that if the body text entering the tube falls below eight hundred characters, the model should never be called and an alarm should sound. The system had no name for an error state. Where a pipeline has no error state, everything succeeds.

There is a second, less discussed truth. When Stage-2 wrote that no deep analysis was possible from empty fields and that it would not invent one, it did what a referee at the monitor is supposed to do — name the clear and obvious error. A referee who cannot name it cannot overturn. This one named it.

Desks run on reflex timing: reacting to a sacking within the hour is a speed that substitutes for credibility. My credibility is built in the delay, not the pace. On Thursday the pipeline chose zero instead of delay. Nothing was fabricated, which is good news. But nobody fabricating anything does not mean there was anything on the desk.

I went looking for a villain and found a filing system instead. That is not better. It is merely closer.

Forward projection: the change that would alter the projection

Three things would be enough, and none of them is expensive. First, an ingest gate: if the body text is under eight hundred characters, the call never reaches the model — hard fail. Second, log prompt length on every call, and the number of parsed keys in every response. Third, the fallback branch should throw rather than return the skeleton. Returning the skeleton is what a referee would be doing if he filed a match report with the teams listed as "home" and "away."

If those three land before the next tournament cycle, the effect will show not in five seasons but in the closing sprint of that one — precisely where the truncation bug lives, where live load is highest, and where the last three paragraphs of a story quietly disappear.

It is the same rule I wrote into that spreadsheet. Do not read an empty cell as zero. The profession, rather than my ledger, is now due to apply it.