The Empty Schema Problem: When a Football News AI Pipeline Fails Silently

Football newsrooms have quietly changed in the past few years. Where a reporter once watched the match, took notes and wrote the story, many organisations now begin with an automated pipeline that pulls information from a web page, fills it into a fixed structure, and hands it to an editor or analyst. On paper the structure looks immaculate: title, source, language, article type, information points, core viewpoints. But when that structure comes back empty, the newsroom lands in an awkward place: work has stopped, yet there is no error message. The system did not crash. It simply filled in nothing. That is precisely what surfaced recently in Stage-1 of a football content pipeline. The output was valid JSON, the schema was correct, the fields were present — but the data inside was missing. Title, source and article type all read N/A. The one-sentence summary was blank. The author's stance and purpose were N/A. The entire information-points section was empty. Viewed from above, the report looked complete; viewed from inside, it was a shell. The most striking part was the supplementary notes. Three lines appeared there, and none of them was data — all three were to-do instructions left for the downstream system. One said: identify from the information points above. Another said: judge from the source fields. In other words, the pipeline had assigned the job to itself, because it had nothing to fill in. It resembles a student who writes the question on the answer sheet and leaves the answer blank. In that entire output only one meaningful signal survived: Domain Label, containing the single word football. The system knew the subject was football and nothing further. This is a dangerous place for any editorial workflow, because a correct domain label creates the illusion that everything else is fine. No red light flashes on the dashboard. Monitoring reports the job as successful. And if the next stage's model is even slightly generous, it will construct a whole narrative on top of that emptiness. That is why refusing to perform a deep analysis on this input is the correct stance. Had an analyst proceeded, whatever was written would have come from imagination, not from source material. In football this risk is especially heavy, because readers treat numbers and names as facts — scorelines, injury durations, transfer fees, contract lengths. A single fabricated detail is enough to damage an outlet's credibility. The right move, therefore, is not analysis but fault localisation. One point must be made explicit: the problem lies in the extraction stage, not the analysis stage. In the football domain the common causes of this kind of failure are limited and can be checked in order of priority. First, crawl or scrape failure — the HTML reaching Stage-1 is near-zero length, or is a login page or paywall. Second, a page that is not static — JavaScript-rendered content, live blogs, real-time score streams. Third, content that is not text — video highlights, photo galleries, embedded tweets. Fourth, a parser that misses its target — the body selector landing on an empty node. Fifth, model degeneration — normal input, yet an empty schema on output. These five causes must be understood separately, because their remedies are entirely different. A crawl failure means the system never got the page, or got something that was not the article. On paywalled sites this is most common: the server sends a teaser rather than the full text, and the pipeline should receive a clear signal that content is unavailable, not a silently empty schema. JavaScript pages present a different problem: the HTML from the server does not contain the article at all, because it is built later in the browser. Live blogs are harder still, since dozens of entries share one page and their order keeps changing. The content-type problem is separate. A great deal of football information is not text. Goal clips, press-conference video, a player's Instagram post, a journalist's update on social media — a text-based pipeline cannot handle any of these. They require speech-to-text or OCR capability, and where that is absent, the honest response is to say the article is not suitable for this process. The dishonest response is to send an empty schema and imply all is well. Parser failure is the most deceptive, because both the page and the server are fine. The full article text exists in the HTML, but the selector lands on a sidebar, a comment section or an empty div. The result is the same emptiness. The remedy is to dump the DOM once and inspect the actual structure. Guessing without looking is merely wasted time. The fifth cause, model degeneration, is the rarest but the hardest to catch. The input is entirely normal and the output looks valid, yet contains nothing. Often it is triggered by changes in signal length, language or format. The solution here is not technical but control-based: enforce a hard check that the number of information points is at least one, and retry when it is not. Across this whole investigation there is exactly one decisive fork: what did Stage-1 actually receive as input. If the input was empty or garbage, the matter is crawling and parsing. If the input contained clean text, the matter is the prompt or output parsing. The two repairs are entirely different, so diagnosis must come first. The cheapest test is to print the input length and the first two hundred characters. To move the pipeline forward, at least one of the minimum requirements must be met. The most reliable option is the clean text of the original article. The second is a URL, in which case it must first be determined whether the text is retrievable, and if not, why. The third is text from a screenshot or PDF. Proceeding without any of these is reckless. At the same time, Stage-1's output must be filled in at least enough to work with: title, source (including outlet, link, author and publication time), language, article type, and the list of information points. Each information point should record where it came from and at what time. Alongside that, a summary of core viewpoints, stance and purpose. The structure is not complex, but claiming analysis without it is a deception. In the football domain the most important column is source_field. It records where the information came from: official, club channel, tier-one journalist, aggregator or anonymous source. In transfer and injury reporting, credibility is almost entirely determined by this column. If it is empty, everything downstream is merely guesswork. The same sentence — the club is interested in the player — becomes three different truths depending on whether it rests on an official statement, a journalist's claim or an agent's briefing. This gap is most visible in transfer reporting. Contract length, wage structure, registration slots and feasibility under financial rules cannot be assessed without the source tier. An anonymous agent's claim and an official club statement do not carry the same weight. Likewise in injury reporting: accurate medical terminology, the gap between official and journalistic claims, and the return timeline — none of these can be checked without the source tier. Once material arrives, the analysis framework runs in four layers. The first is the factual layer, where every information point carries a source grade and conflicting items are placed side by side rather than forcibly reconciled. The second is the entity layer, examining the relationships and interests among club, player, coach, agent and contract. The third is the temporal layer, separating event time, publication time and the present, and flagging which claims were later disproved. The fourth is the argumentative layer, tracing the author's conclusion back through the evidence chain to find causal leaps, sampling bias and problems of data calibration. By article type, additional checks attach. Match reports require verification of score, line-up and key events, and a test of whether expected-goals-type data and the narrative agree. Transfers require the source tier, contract duration and wages, and feasibility under registration and financial rules. Tactical pieces require scrutiny of shape, pressing triggers and build-up structure, and whether the data genuinely supports the conclusion. Injury stories require accuracy of medical language, the gap between official and journalistic accounts, and the return timeline. Opinion pieces require the author's stance, their conflicts of interest, and the strongest counter-argument. At the pipeline level the lesson is clear. Every stage needs strict minimum conditions, and output that fails them must not be counted as success. An empty list is a failure, not a success. Observability is equally necessary — input length, volume of parsed text, number of identified information points, and the reasons articles were rejected. Without these metrics, a fault can stay invisible for weeks while empty schemas quietly move downstream every day. At the industry level the problem is set to grow, not shrink. Football coverage is moving further behind closed doors, live coverage is accelerating, and clubs are pushing information through their own channels. All three mean that clean, static, full text will be harder to obtain. The advantage will go to organisations that treat failure as normal rather than exceptional, and that can state in writing, for every article, where the information came from and how firm it is. The summary of this episode is that valid JSON is not a fact. A correct structure does not guarantee correct content. When a report is filled with N/A and followed by instructions for downstream, it is not a report but a to-do list. Rather than writing analysis on top of it, the more useful question is what the input actually contained. That single question is what can carry a pipeline from an empty schema to a real story.