Stage-1 Came Back Empty: Where Information Dies in the Four-Phase Cricket Coverage Pipeline

3:40 a.m., Mumbai. One tab is open on the desk — the Stage-1 extraction output. The match had ended at 11:20 p.m. At 1:00 a.m. the aggregation layer had pulled forty-one items. At 3:15 a.m. the extractor returned its answer, and the answer was that there was no answer.

Article Title: N/A. Article Source: N/A. Article Type: Unclassified. One-sentence Summary: blank. Author Stance: N/A. Article Purpose: N/A. Information Points: empty list. Entities Involved: "identify from the information points above." Time Sensitivity: unevaluated. Source Quality: "judge from the source fields of the information points."

This is not a case of thin information. It is a case of zero. The match in the middle of it had 240 legal deliveries, two innings, a toss, a DLS check, eleven wicket events. The pipeline reported: nothing. The information had not disappeared. It had failed to pass through a schema.

The method of this piece needs stating in the first hundred words, because what follows is not a match analysis. It is a four-phase pipeline ledger — ingest, parse, extract, route — and I am labelling it openly low-confidence. Anything built on an empty input will be low-confidence by construction. The one field that survived was Domain Label: cricket_world. That is not evidence. It is a routing tag, a hypothesis about the document, not testimony from it.

Context: the ninety-minute window and the two-storey architecture

The Indian cricket digital economy runs on a brutally tight clock. Within ninety minutes of the last ball, dozens of outlets need post-match material — Hindi, English, regional languages, a different angle for each. The BCCI calendar is dense enough that a busy week produces four results and three or four pieces per result. The desk does not first ask whether something is true. The desk first asks whether something is filed.

That pressure produced a two-storey architecture. Stage-1 is extraction — title, source, type, one-sentence summary, author stance, purpose, information points, entities, time sensitivity, source quality. Stage-2 is analysis. Stage-2 can never be better than Stage-1, any more than a review of the sixth over's bowling can exist without the first over's scorecard. No information points means no entity graph, which means no disambiguation, which means no verification path survives.

One distinction gets blurred in most pipeline discussions: a domain label is not evidence, it is a hypothesis about the document. Saying cricket_world does not mean the document is about cricket. It means the pipeline decided to send it to the cricket desk. If the document was in fact about a broadcast deal, or a betting case, the cricket desk receives something it can neither verify nor dismiss. In this trade, being unable to dismiss is more dangerous than being unable to verify.

I have a benchmark, because I have watched a full Stage-1 work. October 24, 2026, Dubai — India 151/7, Pakistan 152/0, a ten-wicket win, Shaheen Shah Afridi 3/31. That night I built a matchup matrix off twenty-four balls: three lengths, three batters — Rohit Sharma, KL Rahul, Virat Kohli — and the outcome for each pairing. The information points that night numbered fourteen. Fourteen points produced eighteen hundred words. Tonight there are zero in place of fourteen, and words do not come out of zero. They get filled in. That is the whole difference.

Core: four phases, and a different death in each

I break a pipeline the way I break a match — into phases, each with its own budget, each budget reconciled. An empty output is not an event. It is a sequence.

Ingest. The first question is not what the extractor said. The first question is whether the raw file exists. If the source HTML never arrived — a 403, a paywall, a changed live-blog endpoint, a scorecard sitting behind a JavaScript widget — then every downstream field is empty by construction. This is the cheapest failure and the one most often hunted in the wrong place. People blame Stage-2 when nobody put fuel in the car.

Parse. If the raw file exists, the emptiness means something else: schema mismatch. The parser wanted headline plus dateline plus body. It got a live blog with an over-comment every three minutes, a video caption of three sentences, or a hundred-word flash with a score and no adjectives. The parser returned nothing because it recognised nothing. An empty output is an ambiguous state — it does not distinguish between "there was no article" and "there was an article and it did not fit my frame." That ambiguity is the expensive part, because in the first case the correct action is silence and in the second it is changing the schema. Both look identical on the output side.

Extract. This is where the real loss happens. The minimum viable unit of an information point is a triple: a statement, its source or speaker, and its time. Lose one of the three and the point does not exist. Statement without source is rumour. Source without time is an old story travelling as new. Time without statement is a calendar entry, not journalism.

Which is why entity extraction cannot proceed from an empty list. The Entities Involved field said "identify from the information points above." There was nothing above. That is a circular dependency, and a logical trap: the pipeline is petitioning its own empty pocket. Same with Source Quality — "judge from the source fields of the information points" — when the information points had no source fields. The result is that no credibility assessment is possible, and without source quality, time sensitivity cannot be assessed either, because time sensitivity is a property of the event, not of the source.

Disambiguation belongs here too, and it is less technical than it sounds. Indian cricket has two players sharing a surname, three abbreviations for one tournament, and two stadiums in one city that have changed names. In a filled list all of that collides and resolves. In an empty list it stays unresolved, and an unresolved entity sits in every subsequent sentence as a possible error.

Route. One field survived — cricket_world. This is where the pipeline does politics. The label decides which desk receives the document, and the desk decides what kind of piece gets made — match report, transfer-market discussion, or regulatory story. When no other field survives to cross-check the label, the label is assumed correct by default. That is the biggest risk in the whole chain. A guess that survives alone gets promoted to evidence.

I connect this sequence to an example from my own desk — May 21, 2026, Hyderabad. Mumbai Indians 129/8, Rising Pune Supergiant 128/8, a one-run margin. The story in the headlines was the final over. The story in the ball-by-ball phase map was eleven dot balls between overs 16 and 18. The phase map said over fourteen; the match said otherwise. That is not a different opinion, it is a different class of evidence — and it came from a log of eleven dot balls, not from a sentence in a report. Had Stage-1 come back empty that night, the whole ledger would never have existed. The final-over story would have survived, because it can be written without any schema at all.

The opposing case: "pen and eyes are enough"

Now the argument against me, in its strongest form, because in its strongest form it might be right.

It runs like this. A reporter in the Wankhede press box does not need a schema. He watched the match. He heard which over was heavy. He has eyes, legs and ground access. The pipeline is a crutch, and the correct response to an empty output is to throw the crutch away and watch the cricket. A system that returned nine empty fields should not be written on at all.

I agree with half of that, and the agreement stops exactly where it assumes an empty output is always a bug. Sometimes it is the right answer. A match washed out after twelve balls, a series that ended without a toss, a DLS-decided result from which no cricket structure survived — in those cases zero information points is an honest return. The correct product there is a 150-word note, not a phase grid. Where fewer than five extractable factual units survive, writing 2,300 words is not analysis, it is filler. I apply that rule to my own model, because the hammer that sees only nails has one signature failure: it decides that anything can be a nail.

But that was not this night. Here 240 balls existed. The correct diagnosis is not "there was no information." The correct diagnosis is "the information was not extracted." Those are different by an order of magnitude, because the first is a fault of nature and the second is a fault of your own tool.

Larger still, and this is the real blind spot: a pipeline failure is invisible in the published product. A piece written on an empty input and a piece written on a full one read the same. The prose does not degrade. The sentences do not break. What degrades is the ratio of verified to inferred sentences — and that ratio is never visible to the reader, because it carries no HTML tag. This is why I put a version tag and the model's assumptions in the first hundred words. It is not etiquette. It is a declaration of how much of the piece is verified, so that when someone later asks where a sentence came from, the answer is not "the pipeline."

Takeaway: a dry run before the next match

This exercise pays only if it becomes a metric. Before the next India fixture, run a dry-run on the last one: count the information points Stage-1 returns. Under five and the product is a note. Above fifteen and the product can carry a phase grid. In between, the product has to state its own limits. And if the count is zero while the scorecard holds 240 balls, fix the pipeline before writing the piece — because what came back empty that night was not information. It was trust.

At 3:47 a.m. I did not close the tab. I left it open and wrote one line above it: check fetch, check schema, then write. Next morning, when someone asks where a number came from, the answer should be the name of a file. Not a sentence.