The Blank cricket_asia Schema: Sixteen Fields, Fifteen N/A, and a Pipeline Post-Mortem

At 11:40 pm the pipeline answered. A JSON opened on screen with sixteen fields, and fifteen of them held an N/A — title N/A, source N/A, article type "unclassified", one-sentence summary empty, author stance N/A, purpose N/A, information points zero, entities "not identifiable", time sensitivity "not assessed", source quality "cannot be judged". One field out of sixteen came back alive: domain_label: cricket_asia.

The four-person desk sits above a mobile-phone shop in Karol Bagh, and its entire job is to turn a cricket story into something that can be re-audited. That night we had a record that existed technically and was empty substantively. This piece is the post-mortem of that record, because a model's misses teach more than its hits.

I went to Kazan to watch a model fail in public. It was instructive, and not for the reason I expected. A model's most expensive output is never the wrong answer; the most expensive output is the empty answer the system quietly books as a success. That is what happened here — not against Germany, but against our own pipeline.

Stage-1 extraction deserves a plain definition. Once a cricket article publishes, a program reads it and breaks it into parts: title, source, type, a one-sentence summary, author stance, purpose, information points, named entities, time sensitivity and source quality. That layer is the raw material for everything downstream — archive, search, context checks, answer capsules, cross-referencing against outside databases. If the layer returns empty, every conclusion built above it becomes guesswork.

The cricket_asia tag makes the job harder rather than easier. Under that one label you can find the Asia Cup, a bilateral series, ICC and ACC governance, board politics, selection and rankings arguments, broadcast rights and league economics, or stadium and grassroots infrastructure. The label is correct and still useless, because a label tells you a subject, not an evidence base.

Here is the audit, field by field, because that is where the damage becomes visible.

Title and source both returning N/A is the largest loss, because without those two fields the article cannot even be located again. No title means no identity; no source means the reliability question never gets asked. All we have left is a sentence — cricket, Asia — and inference for the rest.

A type of "unclassified" means the distinction between news, analysis, opinion, feature and match report has collapsed. That is not a small thing. A claim you accept from a match report is not a claim you accept from an opinion piece. When the type disappears, the reader's confidence level disappears with it, and the desk loses any way to decide what weight this record should carry.

An empty one-sentence summary is the cleanest signal that this is an extraction miss, not an empty article. Summaries come back empty when the article body never reached the program. A paywall, truncated HTML, PDF parsing, encoding or language — one link broke and every field below it was orphaned.

I have spent eighteen years watching cricket and reading its coverage, and I still struggle with the habit of the commentary box: it produces a verdict in one sentence and the scorecard rejects it three overs later. The same gap sits here. Stance and purpose both reading N/A makes any bias analysis impossible. We cannot say whether the writer was neutral or advocating, whether the piece informed or persuaded. Bias is only measurable when a text exists.

Zero information points is the heaviest loss, because that is where the evidence chain begins. Zero points means zero facts, zero quotations, zero numbers, zero causal claims. What survives in this state is commentary, not analysis.

Entities coming back "not identifiable" is honest and dangerous at once. Honest because the system did not lie; dangerous because next time the same system could have quietly filled in a name. This is where the most useful column in my file earns its place — I keep a column for luck. It is the most honest column in the file, because it records whether a failure came from the input or from the inference. In this case that column is empty too, because no input log was kept.

Time sensitivity reading "not assessed" costs most on a tag like cricket_asia. Tournament cycles, governing-body decisions, player availability, rankings and bilateral schedules are all date-bound here. A story one day old in this space is not stale, it is wrong. A record with no date and no time label sits in the archive as a trap.

Source quality "cannot be judged" is a sentence no data desk should tolerate. Without source, author, date and publication context there is no reliability arithmetic to perform.

The diagnosis is clean, and it lives inside the pipeline rather than inside the article. This is an upstream extraction failure, not an article with nothing in it. The candidate causes are finite and all testable: the body was empty, truncated, paywalled or never delivered to the extractor; the parser fell over on HTML, PDF, encoding or language; the prompt and schema returned empty arrays and default N/As; domain classification succeeded while every downstream step was skipped; and JSON validation did not treat an empty mandatory field as an error.

The last cause is the most embarrassing and the most common. The schema treated N/A as a valid value. The pipeline had room to say "I don't know" but no rule to record "I don't know" as a fault.

A real analysis needs at minimum: title, source, author, date and type; a one-sentence summary and stance; five to ten information points each with a source field; named countries, boards, players, tournaments and institutions; time-sensitive facts with dates; and source-quality indicators. Remove any of these and analysis becomes a guessing narrative.

The room's view is that the article simply had no content. I price that view as evidence rather than noise, because here the room is saying something the data cannot directly refute. But the room and the data are stuck at the same point: we do not know whether the article was empty or whether the extractor never received it. Without separating those two, any conclusion is exactly what a par score was in Karol Bagh — in Karol Bagh, the par-score was a rumor repeated until it became a number.

The genuine warning here is the classifier's success, not its failure. The domain label arrived correct while nothing underneath it existed. That is the exact architecture of borrowed credibility: one confident label, zero substance, and then the label gets quoted as a fact for three years. Fifteen empty fields out of sixteen, and the domain label quietly got counted as coverage. Which is why the fix is a contract change rather than a parser change: if a mandatory field is empty, the record is rejected, not half-accepted.

The work ahead is plain — re-run extraction against the raw text, reject records with empty mandatory fields, and log every miss publicly. A monk does not argue with the market. He re-runs the regression. The question is this: of the thousands of records already sitting in our archive that we currently call coverage, how many are blank schemas that nobody has ever reopened?