A Furniture Fair Tagged cricket_asia: Case File of a Category Error in the Data Pipeline

Six Information Points, and One Line That Should Not Have Been There

Eleven in the morning on my desk in Pune, and I had six information points in front of me. One: a furniture fair has opened with a range of designs. Two: several furniture manufacturers are participating. Three: companies are displaying furniture they produced. Four: the fair runs for five days. Five: the venue is International Convention City Bashundhara. Six: the report is a photo story carrying ten images.

And above all of that, the one line that has been the first filter of my working life for fourteen years — the domain label: cricket_asia.

Beneath that line there was no score, no over, no innings, no ball, no batter, no bowler, no team, no league. There was a convention centre, which is not a cricket ground. This is the moment when a data journalist has to choose whether to write a piece or open the ledger. I opened the ledger, because the biggest story in this file is not furniture.

Forty-two years ago I got the same shock in the opposite direction. March 2026, a production truck in Mumbai, and I hand-coded the World Cup final at the Wanderers — Australia 359/2, India 234. The panel said India collapsed. My sheet said India scored at 4.4 an over between the 15th and the 40th, against Australia's 6.1. That day the numbers existed and the words were wrong. Today the words exist, as a label, and the numbers do not exist at all.

Context: What a Label Is, and How I Count It

A glossary is not decoration; it is the spine of an argument. So, definitions first, counting second.

A domain label is the category mark that routes a file to an analyst pool. cricket_asia should mean content played in Asia, or tied to Asian cricket boards, leagues and players. It is a class, not an opinion.

Precision is the share of routed files that landed in the right class. Recall is the share of genuine cricket files that got caught. Mislabel rate is wrong-class files over total files routed. Contamination is the damage a wrong file does once it reaches a downstream dataset.

Stop at the third definition, because it is this report's subject. A mislabel rate needs a denominator. Right now I have one file. One. A denominator of one is useless for any rate. And the journalist who builds a rate without a denominator hands me the commentary-box sentence I learned to hate after 2026 — "the momentum shifted." Which over, against which batter, on which bowler's line, against which field? No answer. No number. So I do not write the sentence.

It matters that this audit lands mid regular season, the window when ingestion volume peaks. Domestic tournaments, franchise media releases, regional photo stories, sponsor announcements — all of it drops into one pipe. The regular season is the season the pipeline is fullest, and in a full pipeline false positives scale by multiplication, not addition. A wrong file seen once in January will be seen eleven times in April.

In 2026 I imposed a forty-term glossary on a new digital desk and banned the word momentum from headlines unless it carried a number. Editors called it rigidity. I call it reproducibility. Apply that rule here and the first question is: what is cricket_asia a denominator of?

Core: Six Analytical Slots, Six N/A, and That Is the Finding

I ran the Stage-1 file through the six slots I open every series report with. All six came back not empty but marked N/A — insufficient information. That is not the failure of analysis; it is the output of analysis.

Slot one, format and match. No Test, ODI, T20 or Hundred. No powerplay, middle overs, death overs or new ball. No score, innings or result. The venue is a convention centre. Weather, dew and DLS revision are not questions here — when there is no match, dew is dew of what? Computing any dew coefficient from this file would repeat what many did in 2026: run a model with no variable in it and call the result a neutral-venue finding.

Slot two, player and performance. Average, strike rate, economy, situational splits, recent trend — all five absent. A category error needs naming here, because it is the root of the whole mess: the names in points two and three are companies, not players. Treating a furniture manufacturer as a player and computing a participation rate is the same methodological crime my trade commits daily when it mistakes a sponsor release for team news.

Slot three, team and ranking. No national side, no sub-national side, no franchise. No ICC ranking, no WTC position, no bilateral record. Batting depth, bowling combination, bench depth, age structure — all four unassessable, because the subject of assessment does not exist.

Slot four, league and commercial ecosystem. No IPL, no BBL, no Hundred. No broadcast value, franchise valuation, salary or auction hammer. One caution is due, because size of economics is a trap: a furniture fair has stall rents, sponsorships and exhibition economics, and none of that is cricket commerce. Every fair has money in it; treating every pot of money as cricket money is a category error. And to anyone arguing that the fair is business and cricket is business and therefore the two share a class — a football transfer market is a spreadsheet with a heartbeat and paperwork, but a wrong row dropped into that spreadsheet stays in it; dressing-room chemistry does not fix it.

Slot five, rules and governance. No ICC, BCCI, ECB or CA. No DRS, anti-corruption provision, eligibility question or geopolitics. Empty for cricket. For the pipeline it logs a medium risk, because the real administrative question is not who won but how this file arrived.

Slot six, risk. No injury, fixture congestion or player loss. The only live risk is data quality: a mislabelled file entering a cricket dataset and generating a false signal.

How did the misroute happen? I will not announce a mechanism without evidence, so three possibilities at three confidence levels. One, acronym collision: International Convention City Bashundhara shortens to ICCB, and cricket's institutional alphabet lives next door to ICC. For a machine the distance between ICC and ICCB is one letter; for an editor it is a continent. (Confidence: medium.) Two, the double life of a place-name: Bashundhara carries both corporate sponsorship and cricket history in this subcontinent, and a keyword classifier does not read that name, it just reacts to it fast. (Confidence: high.) Three, format pattern: "21st National ... 2026" resembles a sports-event announcement, and "National" alone signals a national team. (Confidence: medium.)

All three point to the same place: the error is not in the content but in how the content was read. First define the term, then count the event, then tell the story. There is nothing countable in this file, so the story is that nothing is countable.

When I write N/A in a spreadsheet, it is not an empty cell. It is a logged entry, with a date, a reason, and the name of the person who entered it. An empty cell has no answer; an N/A has a written answer explaining why there is no answer. My whole practice rests on that gap.

Contrarian: One Error Is Not a System; a System Is What Counts the Error

Here I have to break my own most convenient story. The convenient story is: see, automation cannot be trusted. I will not say it, because I lack the denominator.

I have one file and one wrong label. From that I cannot compute a mislabel rate, and I cannot even claim the classifier is broken — a broken system and a blunt keyword filter are different things; the second is doing its job, and its job is sifting rather than thinking. The numbers didn't lie; they never spoke, because there were no numbers there. Where there are no numbers I will not issue a verdict, however attractive the verdict looks.

Second, I am tempted to turn this file into evidence of a larger crisis. No. Correlation is not causation. It is also possible the file was deliberately planted as a pipeline test, in which case its capture is a pass certificate rather than a failure. I will keep both possibilities open because closing them would take evidence I do not have.

The third point is the uncomfortable one. In September 2026 the IPL moved to the UAE and my eleven-season home-advantage coefficient collapsed inside a week — from 0.32 runs per over to 0.06. I killed the variable, rebuilt every venue baseline across sixty matches in nine days, and shipped a neutral-venue model to two franchise analysts. One coefficient I refused to touch: Sharjah's toss-and-dew factor. I will not move it on three weeks of noise. That day I decided to publish my error bars rather than hide them.

The same discipline applies here. If I extract "the pipeline is broken" from one wrong label, I am doing to the classifier what I refused to do to Sharjah — moving a long-sample coefficient on three weeks of noise. Calibrated rigidity means fast on new variables, immovable on long-sample ones.

Fourth, the institutional temptation. I have to admit that if this same label had arrived attached to a board official's or a franchise's press note, the print-desk part of me would have wanted to treat it as the official record. I am naming that temptation because what is not named enters decisions anyway. Every file of mine carries an on-record/off-record ledger. This one has nothing off-record, because no source was spoken to — and that silence is a cited datum, not a gap.

Fifth and last. The greatest risk in this piece is not a wrong conclusion. It is that I make the mistake about myself. At sixty, with a byline, that temptation is real. So the words I have spent on myself here match the words I have spent on the fact, and the correction sits below in one line. The case does not deserve more room than that.

Takeaway: Next Time Ask for the Rejection Rate, Not the Score

Three signals I will watch, all of them larger than one file.

One, publish the rejection rate. Every classification pipeline should state how many files it returned, on what basis, and who looked. A desk that does not publish its rejections cannot claim its acceptances.

Two, a mandatory title-to-label mismatch check. A cheap rule suffices: if the headline contains no sport word and the information points contain no score, player, team or format, the cricket_asia label should stop itself. That is not intelligence, it is procedure — and procedure is cheap.

Three, store N/A as data. Our dashboards currently count only what arrived. They do not count what did not, or why not. The day a dashboard carries a column for "unavailable" is the day we tell fewer stories and more truth.

I trust the ledger more when I know who wrote the entries. The most important entry in this file was written by a machine and read by no one. The question is not what the machine knew. The question is why none of the seventy people downstream read the line — and whether next time we will read it, or skip it in a hurry to see the score.

---

Correction log — entry 47, February 2026: In my first pass I accepted this file's domain label as "probably correct, category broad." Wrong. The content is not cricket, the label is cricket's, and my first look failed to catch the difference. No rate was computed from one file, and none will be.