The Empty Extraction: When Three Hundred Balls Are Logged and Not a Single Run Emerges

On a Monday afternoon at a digital sports desk in Pune, a Stage-1 extraction pipeline returned its output and there was nothing in it. Article title — N/A. Source — N/A. Article type — unclassified. Core viewpoints — empty. The list of information points — empty. And because entities are built from those information points, the entities were empty too. One signal survived the entire system: a single label, cricket_world.

I watched the screen for a while, because this output did not feel new to me. In 2026, tracing a broadcast graphic back to the ball-by-ball log, I found eleven overs in which every delivery was logged as a dot while the scoreboard for the same passage read 78/2. The scorebook was not lying. The sheet was missing. An empty output is not a story; it is the fingerprint of a protocol, and it should be read the way an auditor reads an incomplete ledger — not to find the culprit, but to establish which verification step never existed at all.

Stage-1 has a plain job: pull factual units out of the source text — who, when, where, how much, and from which source. Those information points are the bricks on which the entity graph, the timeline, the head-to-head record and every later conclusion stand. The moment that layer returns empty, everything above it — argument testing, impact projection, fact-checking, chronology — goes hollow on its own. One truth survives for me: the subject is cricket. That is not a small amount of information. It is simply not enough.

My first lesson came from the scorers' sheet, not from a screen graphic. The walk from the Ranji press box to the digital desk taught me one discipline: any index, pressure metric or expected-value model should be reconstructable from the primary scorecard. A number that cannot be traced back to a ball is a rumour with a decimal point. That same test now applies to a stats pipeline. Even one extracted information point should travel back to its ball — meaning its original sentence, its publication source, and its date.

At the 2026 IPL final I applied that rule to myself. A defence of 129/8, fourteen dot balls in the last five overs, and two days of missing sleep at the desk. I rebuilt the pressure index from the scorers' sheet — required rate, wicket probability and bowler-batter matchups compiled separately. I refused to publish a single win-probability figure until the full layer of sixty matches had been checked. The editor asked why the delay; my answer was that I was not being slow, I was building a sample.

The following year the test turned against me. From sixty-four matches of the Russia World Cup I rebuilt an xG model, and it limped in two places — it overrated long-range shots and underrated the pressure generated by set pieces. France's 4-2 final win was a clean test case: the own goal and penalty sequence could not be explained by the model's shot-quality frame. I wrote it that day and still repeat it: the xG audit failed not in the model but in our validation protocol. The sample was not wrong, the control was missing, and the out-of-sample test was never written.

Empty stadiums in 2026 repeated the same lesson. Sixty matches in the UAE, the 2026-versus-2026 venue split, toss effects and dew. Removing the crowd did not kill the toss advantage; it moved the timing of scoring. What would have been casually called a neutral tournament I named only after separating three controls — venue, schedule and team strength. Empty stadiums taught me to separate home advantage from ritual, and the same discipline applies to any system's empty output: fix the controls first, then assign the meaning.

Now put what the empty output contains into technical language. There are two failure modes and their remedies differ. The first: the input was empty, meaning the source article's text never reached Stage-1 and only a URL or headline went in. The second: the text arrived but the extraction output was truncated or failed to parse. The cheapest way to separate them is to look at the shape of the output. N/A in every field and empty brackets where lists should be is not the signature of a model's decision; it is the signature of a parser's fallback.

The scoring analogy is clean. If a log of three hundred balls has every entry marked as a dot, you do not conclude that the innings produced zero runs. You conclude the scoring terminal was disconnected and the sheet should be pulled again. Analytics deserves the same discipline. Reading an empty output as a zero value is the most expensive mistake available, because zero is visible on screen and the absence of data is not. An empty field starts behaving like a finding, and soon an editor writes the story that never existed.

The reflex will be to blame the model. I do not take that road. The audit-bound sceptic places a finger on the protocol first, because whenever the post-mortems of failed analytics claims in Indian cricket coverage were written, the validation design turned out to be the weak link — wrong sample, missing control, no out-of-sample test. This case carries a third cause as well: language. If the source material is not in English and the extraction prompt does not demand that language, the pipeline quietly returns empty, with no error message, only N/A. That is not model incapacity; it is prompt engineering oversight.

There is another trap I keep finding in my own work. Once the pipeline is repaired, people immediately build entity graphs and hang large claims on them — who dominates whom, which player rises against which side. A graph built without ball-level provenance is still a rumour with a decimal point. Before a transfer fee I check the ledger of minutes played; before a valuation of any cricketer I check his deliveries — how many, under what pressure, on which line. An auction purse is not a performance metric, and that gets settled on the day you decide which verification door is the first one.

An empty Stage-1 is no threat to the cricket world. The threat is a desk filling it with another layer in a hurry and printing that as fact. My condition is clear: the two-thousand-word analysis above will exist only when three dozen facts below it exist with dates and sources attached. Anything written before that is not analysis; it is decoration applied to a guess.

So when the next desk says the pipeline has been fixed, ask three questions and take the answers in writing: was the input empty or the extraction failed — because the two repairs differ; do the source language and the prompt language match; and when and on how many samples was the out-of-sample test of the new numbers run. A desk that answers all three earns my trust on its ball-by-ball data. A desk that does not leaves the scorebook closed — and with the scorebook closed, the most honest report available is that we do not yet know.