Kaptai Lake, cricket_asia, and One Wrong Label: The Corpus Defect Nobody Is Counting

The record carried ten information points and not one of them was about cricket. The first said tourist numbers had risen along Kaptai Lake in Rangamati district during the Durga Puja holidays. The second said the cottages at Rangadwip Resort were full. The third said Furmon Resort was crowded. The fourth said the same of Wellwood Resort. The fifth described queues on the road to Subalong Waterfall. The sixth had the tourist suspension bridge buzzing. The seventh counted the crowd at Deer Park. The eighth noted rising demand for boat rides. The ninth flagged pressure on the roads because of the holiday. The tenth was the author's opinion that the atmosphere was "alive with the festive holiday spirit."

Ten out of ten were tourism points. No team. No player. No board. No match. No rule. No broadcast deal. And the domain label on this record read: cricket_asia. Ten out of ten also carried a blank source field. No outlet named, no reporter named, no date.

This report is about that one record. Not the lake. The gate it walked through.

I keep a ledger I built by hand. Eleven seasons of Hyderabad first-class scorecards, every dismissal tagged by type — LBW, caught, bowled, stumped, run out, and the residue that fell into "other." No model told me which dismissal was which. I read every line. When the 2026-21 Ranji Trophy was cancelled — the first season not played since 2026-35 — I had scorecards and no ground, and I kept score for a season that never bowled a ball. That exercise left a habit: a thing that is not recorded is as solid to me as a thing that is, provided I can say why it went unrecorded.

That is why this record bothers me. Its problem is not what is missing. It is what got in.

A cricket-Asia corpus is the raw material for three things. A sentiment index — how hot the conversation around Asian cricket ran in a given week. An entity co-occurrence graph — which names keep arriving alongside which other names. A volume-based trend model — what got written about, and when. All three rest on one condition: whatever is inside the corpus has to be cricket. The day that condition breaks, all three lose their definitions.

Now the arithmetic, which I ran repeatedly and which came out the same each time. Take a batch of five thousand records. If two per cent of them contain no cricket entity at all — and two per cent is precisely the threshold at which a sample audit should trip — then one hundred records are bad. Each tourism record carries roughly eight place-names: a lake, three resorts, a waterfall, a bridge, a park, a boat jetty. One hundred times eight is eight hundred spurious co-occurrences. Drop those eight hundred into the same graph where the real names of Bangladesh cricket are floating. Because every false record is stamped cricket_asia, the graph has no reason to quarantine any of them. A lake in the corpus makes a week look quiet in cricket terms, and quiet never names its source.

One clarification matters here, because this is exactly where such reports usually slip through. None of the ten tourism points arrived with a number. "Tourist numbers increased" — increased by how much, against which baseline, sourced to whom. Nothing. The remaining eight points use impressionistic verbs: busy, buzzing, drawing crowds. The tenth is explicitly the author's opinion. The report is descriptive, not analytical. Even as tourism data it is an impression, not a statistic.

My interest is not in the tourism. My interest is in how this record became cricket in the first place.

Two tokens. Just two.

The first is "tourists." In English the same word describes a touring cricket side. Any keyword router that folds "touring side" and "tourists" into one bucket will push this record toward cricket. The second is "Asia." Rangamati is in Asia. Bangladesh is in Asia. The Chittagong Hill Tracts are in Asia. The failure is not that the model guessed wrong. The failure is that the definition of cricket_asia contains no cricket — it contains geography and a keyword. A label with no room inside it to say "not cricket" will always say cricket. This is a taxonomy failure, not a model failure.

There is a third token, and it is the most dangerous because it lives in a calendar rather than a word. The Durga Puja holidays fall in the late-September-to-October window in which Bangladesh's domestic season opens. From one district, along one river, on one road, in the same weeks, two kinds of copy emerge — one from a travel desk, one from a sports desk. The router is not being stupid. It is catching a genuine seasonal confound and calling it a label. The error is not in the words. It is in the calendar. I will not print that sentence until the board's fixture list confirms domestic matches were actually scheduled in that window. That is my rule, and I built it the day a retention list was jumped by nine minutes.

The fourth cost is in the names. Rangadwip, Furmon, Wellwood — three resorts. Entity-resolution systems run on string similarity and co-occurrence priors. A record labelled cricket_asia in which "Bangladesh," "Chittagong" and "Rangamati" appear together hands those three names a prior that drags them inside the cricket economy. Downstream, a franchise-valuation model can read them as new commercial actors in the Bangladesh cricket market. Names that belong to resorts can be read as franchises, and once read they do not come out of the graph. None of the three has any link to a Bangladesh Premier League franchise; the record contains no such signal. But absence is not evidence when the system never asked for evidence.

The fifth and heaviest cost is that the record cannot be withdrawn. All ten points have blank source fields. A record with no outlet, no author and no date cannot be verified, cannot be retracted and cannot be cleanly deleted. It can only be quarantined. I have read matches off a scorebook for ten years on one unbroken rule: no tactical sentence without a clip behind it. There is no clip here. There is a lake and a label.

Here is the part that stings most, and it is not about cricket. The only thing this corpus holds from Rangamati district this cycle is a lake. What it does not hold is a scorecard. I tagged every dismissal across eleven seasons of a ledger because I wanted to see what kind of bowler the Hyderabad pipeline was producing. The answer was white-ball bowlers who could not hold a red-ball length. I did not hear that conclusion. I counted it. By the same logic, if a district produced no scorecard in the corpus across a season while producing a resort report, that is a coverage question, not a data question. The archive nobody reads does not quietly disappear. It gets replaced. Here it was replaced by a lake.

I know the next reaction. Someone will say: fix the model. Someone will say: delete the record and move on. Both are instinctive and both are wrong, for the same reason — both assume the problem lives in the record. It does not. The record was honest about what it was: ten points, all tourism, all unsourced. The lie came from the label.

And here the argument inverts. Do not delete it. Quarantine it.

The reason is plain: this record is the only evidence that the ingestion gate failed. Deleting it deletes the audit trail. If one record in a batch was mislabelled, its siblings from the same feed almost certainly were too — that is an inference, not a finding, and it can be tested only if the sample survives. A record you removed is a record you cannot audit.

The second point will sound stranger still. The tourism data here is not entirely useless to cricket — just not in the way anyone would guess. Occupancy at Rangadwip, Furmon and Wellwood is not an input to franchise valuation. It is a signal of transport and hospitality load in a district during a holiday peak. If a board ever needed to move a team through that district — for an age-group fixture, for a camp — that load peaks in exactly those weeks. This is a hypothesis, not a conclusion. There is one way to test it: open the fixture list for that window. If no match was scheduled in that district, the hypothesis dies and the record is simply waste. A reversal that cannot be tested against a document is not a reversal. It is a habit.

What frightens me here is not that a tourism report entered a cricket corpus. It is that if this record could get in with ten blank source fields, then something else got in through the same gate with its source fields filled. Having a source is not verification. Having a source only tells you whose name the error will be filed under.

Three signals I now watch, and would advise any desk building a sports corpus to watch. First, mislabel rate: sample any batch and count the records containing no team, player, league or board entity. If that figure passes two per cent, corpus quality has already collapsed even while the index still runs. Second, source completeness: the share of information points carrying a named source. The day that ratio drops, the right of retraction goes with it. Third, batch provenance: group records by originating outlet. If several off-topic records arrive from one feed, the problem is feed mapping rather than modelling — and it is the cheapest error available, because it can be fixed at the door.

My ledger had a column for the type of every dismissal, a column for the over, a column for the bowler. The model had a column for everything except whether it was cricket. I filled that column myself, line by line, by hand, because I did not trust anyone else to fill it. That, I think, is where we are now. We built models to compute bowling averages, graphs, indices, and forgot to decide what gets in. Expected goals is a diary written in decimals — but a diary only works when the pages hold the game and not the weather.

Someone is keeping score. The question is who is keeping score for the corpus, and which column they are filing it under.