Empty Field, Green 'OK': An Autopsy of Silent Failure in a Cricket Data Desk

At 1:40 am in Ahmedabad, the pipeline logged stage_1: completed in green. Beneath it the payload sat empty: article_title N/A, one_sentence_summary blank, information_points an empty array, entities never resolved, time_sensitivity and source_quality both N/A. The only valid signal in the entire structure was domain_label: cricket_world — a vocabulary hint, not a fact. I opened the file twice, scrolled a third time, and found nothing to trace. In 2026 I could trace a no-ball forward through a frame map because the sheet contained a ball, an over, a field. That discipline — tape first, arithmetic second, prose last — cannot run on a blank node. Stage-1 feeds at least seven downstream consumers: headline generation, tagging, entity linking, GEO capsules, related-question generation, archive indexing, and the newsroom priority score. An empty information_points node breaks all seven at once, and no step in the chain raises an alarm, because every step is internally correct. A valid N/A in the schema converts failure into success. Empty and unknown are different states and must not share a field. Cricket is the worst possible domain for this stack. Scorecards live on separate pages from the article; live blogs put the most valuable update at the bottom of a 400-entry page; entity resolution breaks on the four names every Indian player carries (English, Devanagari, transliteration, commentary nickname); Devanagari prose packs three fields into one sentence. During retention and auction weeks this matters most, because the money moves through the entity graph, not the prose. The consensus is correct and I concede it: the scraper is broken, fix it. But the consensus hides a false premise — that the empty sheet is a failure to be suppressed. It is the most honest output the system produces, the only run in which the machine did not lie. The dangerous run is the neighbouring one, where a model fills the gap with a plausible name and the green check survives. Volume incentives push exactly that way. With an illustrative six per cent silent-empty rate on a forty-item desk, roughly 870 hollow items a year pass as processed — and that figure carries an explicit confidence band, because one night is not a season. Next time, check raw-text length first, canonical URL pattern second, JSON-repair count third, model output last. Make N/A a hard fail. Then count how many items in the archive finished green and never held a fact.