An Empty Cell Is Also a Measurement: The Silence in Cricket's Data Pipeline Nobody Files a Report On

On Monday morning I opened a report card with thirty-two fields. Twenty-nine said N/A. Two held empty lists. One line survived: Domain Label — cricket_world. I scrolled up and down the way you scroll a scorecard that records the toss and nothing after it. No batters, no bowlers, no overs, no wickets. Only proof that the ground existed.

I have worked with cricket numbers for fifteen years, and an empty sheet does not frighten me. The filled sheet does. An empty sheet indicts itself; it tells you plainly that it saw nothing. A filled sheet makes it almost impossible to separate the number that was measured from the number that was inferred.

The Stage-1 output was a clean null. Title N/A. Source N/A. Article type unclassified. Core viewpoints, stance, purpose — all three blank. The information-points list empty. And entities, which depend on those information points, empty by inheritance.

One signal stayed alive. This report starts from that single line, and the line is itself the finding: when the bottom layer of a pipeline breaks, every layer above it starts to look identical — empty. That symmetry is the real hazard, because it removes any way of reading where the break happened.

Context: Stage-1 is the scorer's book

Cricket keeps two kinds of record. The scorecard, with twelve names, four over-summaries and a result. And the scorer's book, with a line for every delivery — who bowled, who faced, where it went, who collected it, what it cost. Television shows the first. Models run on the second.

Stage-1 extraction is that book. Raw text goes in; four things come out: the writer's central claim, their stance, their purpose, and the list of factual points the claim rests on. Entities fall out of those points — players, teams, tournaments, institutions, dates. Entities build a citation graph. Every later layer — form, auction value, delivery sequence, home advantage — stands on that foundation.

What arrived had no first brick. With title and source both N/A, even provenance had failed. A document whose author and date are both unknown cannot be trusted or distrusted; trust requires knowing what is in front of you.

My own entry into this work began with that book. In 2026, a twenty-two-year-old economics undergraduate at Gujarat University in Ahmedabad, I spent six weeks hand-coding all sixty matches of an IPL season from Cricbuzz ball-by-ball logs — roughly fourteen thousand deliveries, each needing bowler, batter, runs, wicket type, fielding position. I built a sequence-entropy score for Rashid Khan's leg-break against his googly and wrote 2,400 words arguing the wicket came from the wrist, not the pitch. It drew 4,100 reads and one direct message from a three-person analytics start-up in Mumbai.

Those six weeks taught the lesson that applies directly here: extraction is labour, and labour fails quietly. No alarm, no error string. The row that should exist simply does not, and in its place sits a blank that looks tidy.

N/A and an empty list are not the same thing

N/A means the question was asked and no answer came back. An empty list means an answer came back carrying nothing. The first indicts the model or the input. The second may indict the document, which might genuinely have contained no verifiable claim — a preview, a colour piece, a rumour with no date and no attribution. Merging the two is like putting a duck and a not-out in one column.

The severity order is fixed. Title and source at N/A is worst, because it shuts the provenance door. Type unclassified is second, because type determines the analytical frame — news, comment, data report, preview — and each carries a different evidentiary bar. Empty core viewpoints rank third. Empty information points rank fourth, yet they are what the eye catches first, because a blank list is visible and an N/A field is not.

The dependency ladder breaks at the bottom and shows at the top

Entities are empty because information points are empty. True, and incomplete. The real question is not why the list is empty but what the list was supposed to be built from, and where that went. Until that is answered, every layer above stays under suspicion no matter how correct it looks.

A null propagates upward and turns symmetric on the way — every field equally blank. Readers take that symmetry as proof the document was empty.

I have watched this happen inside a franchise review sheet. One column comes back blank — a fitness flag, say — and three days later nobody in the room asks why. The room assumes the report never arrived. Then the player is on the field in the fourteenth over.

The signature of a parse fallback

The shape of the output is itself a fingerprint: valid JSON, every key present, every value blank. When a model's output is truncated or fails to parse, the system usually does not throw. It fills a shell that satisfies the schema and moves on, because a shell is cheaper than an exception.

It is the DRS rule: when the signal fails, the previous decision stands. Nothing is logged as broken; the old state simply persists. Twenty overs later nobody can tell that the technology did not work on that one ball.

So my first check is not why the data is empty. My first check is whether the data ever went in. Three things in order: did the full body text reach Stage-1, or was a URL and a headline passed off as input; was the output truncated at a token ceiling, since truncated and empty look nearly identical; and language.

Language is the variable the prompt forgets

If the source is Hindi or Marathi and the schema demands English output, the model faces three routes — translate, answer in the source language and break the schema, or return a shell that fits. The third looks safest and does the most damage. Entity resolution compounds it: S. Gill, Shubman Gill and शुभमन गिल are three strings and one player, and until they collapse into one node, the citation graph splits a single career into five.

2026: the right model, and the call that never came

Nine months later I was the third employee of that start-up on ₹28,000 a month, building a value-per-crore model for the January 2026 mega auction in Bengaluru. My sheet ranked an eighteen-year-old Shubman Gill above three capped internationals. The franchise I was assigned to bought experience. Kolkata Knight Riders took Gill for ₹1.8 crore, and he made 203 runs in his first season.

My model was right and nobody called. A correct model nobody adopts is an empty cell in the decision layer, and it is the most expensive null I know. I began publishing auction pieces with two columns — model price and market price — and one line naming who was overpaid.

2026: eighty-four days, forty pages, 0.31 to 0.04

In the IPL 2026 bubble in Dubai I spent ninety-four days embedded with Kings XI Punjab. My super-over simulation ranked the captain's chosen bowler fourth. Punjab lost super overs to Delhi Capitals on 20 September and to Mumbai Indians on 18 October and missed the playoffs on net run rate. Across sixty crowdless matches my home-advantage coefficient fell from 0.31 to 0.04.

I said nothing in the review meeting, then wrote a forty-page memo nobody read and posted an anonymised version. The real null in that room was my silence — a field that was asked a question and returned N/A.

Contrarian: a filled JSON, an empty truth

The danger is not a pipeline that returns nothing. It is a pipeline that fills. At schema level, a filled JSON and a correct extraction are indistinguishable, and downstream systems accept both without complaint. An empty output at least declares its own failure.

A second trap sits alongside it, and I know it from my own habits: treating pipeline output as testimony rather than material. Attribute the claim, name the incentive, test it against the record in the same piece — or cut it.

And sometimes the null is the finding. Some documents genuinely contain no verifiable claim. From outside, though, we cannot separate a document's emptiness from a pipeline's. Until we can, an unresolved question sits beneath every conclusion we publish.

Takeaway

The repair order should not be renegotiated: confirm the full body text arrived uncut, check the language instruction, re-run, then cross-check against the source. The final step matters most — publish the null. Log it the way we log a duck. Zero is a number too.

My sheet has carried a column called source status for three years. It should never read N/A. On the day it does, nothing else on the sheet deserves belief.

The question is not whether one pipeline failed. It is how many published numbers we are standing on top of an empty cell that nobody ever reported.