The Empty JSON: The Cricket Coverage Failure That Never Shows an Error

At 2:17 in the morning in a Pune apartment, two monitors were lit and a third screen scrolled a pipeline log by itself. The response arrived — a JSON payload of 1.2 kilobytes. Seven fields. Six marked N/A. In the seventh, two words: cricket_world. No red flag, no alert, no stack trace. Status: success.

The article under review existed somewhere. Someone had written it, a desk had published it, an editor had put a headline on it. The machine extracted not one fact from it. Worse — the machine never said it had found nothing. It said the job was done. In twenty-one years of reading scorecards I had never seriously believed that an empty scorecard is more dangerous than a wrong one. I believe it now.

The gate that closes without a sound

Modern cricket coverage runs on three layers. Crawl the page. Extract the body text. Then Stage-1: force that text into a schema — title, source, article type, one-sentence summary, author stance, information points, entities, time sensitivity, source quality. Stage-1 is the real gate. Everything downstream depends on whether anything came through it. If information points are empty, every later step — claim verification, bias detection, separating fact from opinion — is decoration. You are not auditing anything. You are dressing up a guess.

Cricket is unusually hostile to machine reading. Board and state-association pages are JavaScript-rendered, so static HTML returns nothing. Live blogs hold four hundred updates on a single URL and body-extractors typically grab the last three paragraphs, dropping the other forty overs. Rights-holder pages sit behind paywalls. Auction lists arrive as PDFs whose columns are separate only to the eye. Hindi, Tamil and Telugu pages change the tokeniser, and a model trained on English stumbles on names and counts. In 2026, when I entered digital cricket media, a scorecard was a newspaper clipping read out over a phone. Every layer added since has made the record richer and more brittle in the same motion.

Seven fields, six empty: an X-ray of a failure

The Stage-1 report lay open in front of me and its most honest feature was its emptiness. Title, source and article type — all N/A. The identity of the thing being analysed was unknown. The one-sentence summary was blank, meaning no central proposition existed to take apart. Author stance and purpose were N/A, meaning it was impossible to tell whether the copy was reporting, a column, a press release or franchise marketing.

The entire information-points section contained no entries. This is where the failure turns fatal. Everything else could be inferred around; this could not. No facts, no numbers, no quotes, no events. Entities were unidentified — no subjects, no relationships, no network on which to hang a question. Time sensitivity was never assessed, so nobody knew whether the material was from today or from six months ago, a distinction that decides everything in a sport where a season change turns a true sentence false. Source quality was unrated, so there was no way left to rank evidence by strength.

One signal survived: cricket_world. That label is not a description of an article; it is the name of the failure that deleted the article. It tells me which vocabulary to think in — bowling phases, net run rate, auction purse, rights cycles. It cannot give me a single claim I could print.

Four causes are plausible, and I have seen them often enough to rank them. First, body extraction returned empty: anti-scraping protection, JavaScript rendering, a paywall, or a page carrying only a headline. Second, the URL supplied was a list or navigation page rather than an article. Third, the language model's response was truncated, JSON parsing failed, and the system silently substituted an empty structure. Fourth, and least excusable, nobody looked at the empty fields because the pipeline had reported success.

The disease is not bad data. It is silent degradation. A pipeline that crashes is honest — it makes a noise. A pipeline that returns empty-handed and writes status 200 is lying, and its lie never reaches an alert queue.

The one-run final taught me in 2026 that certainty is a fragile model. In Hyderabad, with eight overs left, my model had us at 61 percent. We lost by one run. Since that night I file nothing within twenty-four hours of a knockout and publish a decision log instead of a reaction column — selection, matchup, field placement — so readers judge the process rather than the scoreboard. Every automated cricket report should carry its extraction-liveness alongside it, or it is simply confidence borrowed against itself.

During the 2026 shutdown I encoded all 757 IPL matches played to that point, ball by ball. What troubled me most in that work was not a wrong score but an incomplete file. A match should contain 240 deliveries; if the file stops at 183 and raises no error, the dataset is not alive, it is sick. Three basic checks would have caught this case.

First, the ratio of raw bytes to cleaned text. A 1.2-kilobyte payload announces that no article was ever inside. Second, the entity count. cricket_world is a category, not an entity. Any meaningful cricket text contains at least one team, player or tournament name; zero entities means zero content. Third, date parsing and source classification. If no date was found and no source type assigned, you do not know what you are trusting.

I have a habit of borrowing from football, and it taught me that data travels better than tactics. At the 2026 World Cup I built an expected-goals model and a pressing index from scratch in a rented flat across sixty-four matches. I came back to Pune and stripped the football out of it — expected runs per ball by phase, control percentage, a pressure index built from dot-ball clusters. Two Indian digital desks rejected the framework outright. The third ran it through India's 4–1 defeat in England, where my top-order false-shot rate read 27 percent inside the first ten overs. That gave me two rules I still follow: attach a one-line definition to every number, and tag anything genuinely borrowed from football, because Indian readers will forgive a foreign idea but never an undefined one.

By the same logic I want a one-line definition on cricket_world. It is a signal for language and prior, not for fact. Treating a category label as a summary is the same error as treating a model's output as prophecy in a one-run final.

The most dangerous place is where there is no data at all

The risk is not that the pipeline produced nothing. The risk is that a label like cricket_world plus a headline is enough for a language model to produce eight hundred confident words about cricket. The model does not need facts. It needs a register. And that is precisely the trap an analyst falls into most easily, because the profession's entire incentive is to be useful. At 2:17 in the morning, with an empty screen and a reader waiting, the pull to fill the void was real. I did not fill it. I will not pretend the pull was not there.

The quieter point is that the pipeline's failure says nothing at all about the quality of the source article. That is correlation against causation: a machine failing to see something is not evidence that the thing was absent or bad. The piece may have been solid reporting stopped by a JavaScript wall, or it may have been a press release. I have no basis to decide. Publishing a 'crisis in cricket coverage' piece on top of that uncertainty would be marketing a guess as journalism.

And because it cuts against my own instincts, I will write it anyway: the most useful output of this failure is that it applies to itself. The pipeline I trust can return empty just as silently as a franchise's rights data, a state association's attendance figures, or a WPL broadcast promise. The things cricket coverage declines to measure — empty stands, regional-language reach, the Ranji Trophy's thin footprint outside knockout weeks — share one property: they make no noise. The empty-stadium study taught me that silence has a measurable weight. The same lesson now applies to the data layer. What goes unmeasured does not disappear. It simply goes missing without saying so.

As a data monk I fast on small samples and feast on validation. This was an odd feast: a null result that proves discipline applied to itself. Zero is also a result. It is just not edible until the reason for it is on the record.

The next-round signal

What I will watch next is not a new statistic but its disclosure. Ask any data desk two questions: what was your extraction-liveness rate last season, and how many null results did you publish? A desk that cannot answer either is selling borrowed certainty. Of all the insights printed on Indian cricket this season, how many came from a payload that should have returned empty — and who decided to fill the silence instead?