From a Ranji Ledger to the IPL Auction: 1,986 Balls of Accounting and a Sealed-Bid Market

Stage-1 came back empty. No title, no source, no list of information points, an entity graph of zero. The only surviving signal was the domain label: cricket_world. I entered that label in my ledger and decided on the spot that this time the method would be the story, not the news. Because when a report contains not the ball but the claim of a ball, the journalist's job is not to repeat the claim but to measure it.

The ledger opened at 1,986 balls, and the season started confessing. It was the 2026-89 Ranji season, Gujarat's campaign, the old Sardar Patel Stadium in Ahmedabad, and the Ahmedabad edition of The Times of India. The press-box habit then was clean: write the wickets and the boundaries, the rest is weather. I dropped the habit and logged every single delivery of six matches in a hardbound book — line, length, field position, shot. 1,986 balls in all. Gujarat finished mid-table; nobody asked for the book. I photocopied the phase-by-phase economy tables and mailed them to the state selectors. In the off-season bowling rotation debate, they used those tables. That was my first lesson: what has not been counted does not survive the argument.

In May 2026 the match-fixing story broke. I was on Cricinfo's India desk. There was no chance I would chase quotes. I pulled ball-by-ball data from 214 ODIs played between 2026 and 2026 and built an "anomaly index" — the variance in run rate and the clustering of wickets in dead rubbers versus live matches. A small set of games could not be explained by form at all. I published the method in full, named no player without documentary proof, and the piece was cited across the inquiry coverage. It also earned me three legal notices. I entered those notices in the same ledger too — one column 'claim', one 'evidence', one 'outcome'.

February 20, 2026, a hotel ballroom in Mumbai. Eight franchises, a US$5 million cap each, MS Dhoni the costliest buy at US$1.5 million. Six weeks earlier I had begun converting Ranji one-day strike rates and domestic T20 footage into a value-per-crore model: runs above replacement, death-overs economy, wicketkeeping value, the age curve. Within minutes of the hammer falling I flagged the Andrew Symonds (US$1.35 million) and Ishant Sharma (US$950,000) prices as overpays. Three franchise analysts asked for the spreadsheet the following week. The auction had a cap; the spreadsheet had a conscience.

Now I apply the same method to an empty output, because an empty output is also a result and it too belongs in the ledger.

Context: a report that carried the name of the ball, not the ball

The material that reached me had N/A for both title and source. Article type unclassified. Core viewpoints — one-line summary, stance, purpose — all three blank. The list of information points empty, and since entities depend on information points, no player, team, tournament or organisation reference graph could be built. This is not an analytical failure; it is a data-entry failure. In football's language, this is the match whose scoresheet was never filled in while we argue about possession.

For a data journalist that difference is the difference between life and death. If the source has no name, I cannot establish who made the claim — a franchise press release, a board bulletin, or an agent's WhatsApp forward. If the article type is undetermined, I do not know whether I am reading news, opinion, a data report, or a preview. All three have different standards. News needs two independent sources; opinion needs the author's track record; a data report needs sample size and window; a preview needs a pitch report and injury updates. Not one of the four is present.

One thing existed — the domain label: cricket_world. It tells us the subject is cricket. That is all. It is as useful as saying a man has something in his pocket without saying how much, in what currency, or whose.

Core: three possible causes of an empty output, and three different checks

Now the part I am writing this piece for. I treat an empty Stage-1 as a disease, and the cure begins with diagnosis. There are three possibilities, and each has a different test.

Possibility one: the input itself was empty. This is the most common and least discussed failure. Very often a URL is sent down the pipeline, not the text. Or a headline is sent and the scraper is assumed to bring the rest. If the scraper stalled on a paywall, a JavaScript render or a region lock, then what Stage-1 received was an empty string. On an empty string no model, however good, returns anything but zero. I call this the 'phantom match' — the scoresheet records a match, nobody came to the ground. The test is direct: look at the length (character count) and the first and last 200 characters of the text handed to Stage-1. If the length is zero or a few hundred characters, the problem is below, not above.

Possibility two: the input was fine but the output failed to parse. This is technically the most deceptive failure, because the model did the work and the result fell off en route. The signature is clear: every field N/A, every list empty, no error message. This is the classic footprint of a parsing fallback — when the JSON parser does not receive valid JSON, it returns the default schema, and the default schema has N/A everywhere. It reminds me of the moment in DRS when the ball-tracking frame drops and the system shows 'undecided' — the camera was not broken, the frame was out of sync. The test: look at the raw model response, before parsing. If text is there and did not fit the schema, the problem is in the prompt format, not the data.

Possibility three: a language collision. If the original article was in Hindi, Gujarati or Urdu and the Stage-1 prompt was English-centric, the model had no clear instruction to extract. This is the same error that happens daily in domestic cricket coverage: Ranji stories are printed in regional languages and never enter the national database. The result — the selection debate names those who have English coverage, not those who have performance. This is the hidden similarity between an empty Stage-1 output and Ranji selection: in both places the information existed, but its language was not recorded.

Beyond these three there is a fourth, less technical and more structural cause: an over-rigid schema. If the Stage-1 schema demands 12 mandatory fields and the model can fill 9, many pipelines reject the entire output — turning partial success into total failure. This is the same error a franchise makes at auction when it adopts an all-or-nothing strategy, saves budget and still leaves the squad incomplete. Partial information is still information; the decision to reject it is also a decision, and it too should be accounted for.

Contrarian: the rush to fill an empty output is the costliest mistake

Here I speak against my own profession's habit. A data journalist's natural instinct is to fill the empty space somehow — because empty space makes the reader feel incomplete, and making the reader feel incomplete makes the publisher unhappy. But that is where the biggest risk hides.

If I fill this empty Stage-1 with an estimated title, source and information points, I will be doing exactly what a franchise does when it buys a player without a scouting report — the team looks complete, and the risk shows up next season. And the reader will never know which fact arrived and which was filled in. That is a breach of transparency, and in my ledger it would be the gravest entry.

So my position is clear: when there is no source, the most honest report is the one that reports the absence of the source. That is not an excuse; it is a result. And that result has value too — it tells us where the pipeline leaks. If I hide it, the leak will never be caught, and the same empty output will come again, and again someone will fill it.

In my experience the most dangerous report is not the one with an error; it is the one with a system for hiding the error. After the 2026 anomaly index I received three legal notices, but nobody questioned my method, because I had printed every variable and every limit. Transparency is the strongest armour — legal as well as professional.

Takeaway: what to count next time

Now the part this piece was really written for. If Stage-1 is to run again next round, I will count four things and enter them in the same ledger that holds the 1,986 balls of 2026-89.

One — the input character count. If it is zero, blaming the model is pointless. Two — the raw response length and its first and last token. If the raw response is full and the parsed output empty, the problem is in the schema. Three — the language tag. Both the original text's language and the prompt's language should be recorded. Four — the accounting of partial success. If 9 of 12 fields are filled, record it as '9/12', not as 'failed'.

This four-column ledger is worth more than my next investment call, because it shows where the information was lost — and the place where information is lost is the place to check first next time.

I audited 214 one-day internationals and found a pattern with fingerprints. This time I audited an empty output and found a pattern — the pattern was not in the tool but in the process around the tool. The auction had a cap; the spreadsheet had a conscience. This time there was no cap, no spreadsheet — only an empty column waiting for someone. The question is not who filled it. The question is who asked, before filling it, why it was empty.