41 Information Points, 28 Opinions, 0 Independent Sources: Inside a Contaminated Football Data Pipeline

At 2:14 in the morning I opened the seventh folder in a batch the newsroom system had tagged "football". Page one read: The Fullton Regal. Page four: a 7.1-metre void. Page nine: a 12-metre frontage. I scrolled to the end of the file, opened the domain field, and read one word: football. Confidence: high.

I went back to page one and read more slowly. Forty-one information points. Not one club. Not one player. No coach, no match, no goal, no contract, no league table, no injury, no suspension. Only twin villas, commercial townhouses, 150 to 180 square metres, 115 to 174 square metres, 75 amenities, two parking spaces, separated stair and lift cores, two theme parks, closed security.

What kept me at my desk near three in the morning was not that I had misread a document. It was that the system had not hesitated.

Four steps and one data field

Sports news runs on four steps: collect, label, summarise, publish. At step two, a single field decides the fate of the entire document. If the label says football, the document goes to the football desk, the football summariser, the football training corpus. If the label is wrong, every layer behind it inherits the error, and no layer has the authority to overrule the one before it.

In nine years of watching this industry I have seen the fault at much smaller scale. A transfer story tagged "basketball" because it contained the phrase "transfer window". An injury piece tagged "medical". Those were harmless, because readers still spotted the error and moved on. This one is different. This document does not travel to readers. It travels to a machine that is being used to write for readers.

And the document does not hide itself. It declares its own type as "product introduction