Method · Language
Evidence and structure come before composition
When the writing comes first, the evidence ends up fitting the text; when it comes last, the text fits the evidence.
A reporting system that starts by generating text inherits a problem it never recovers from: every figure ends up trapped inside a sentence. Correcting a piece of data means rewriting the paragraph, and checking a claim means reading prose instead of comparing values.
We have ordered the process the other way round. First the evidence is gathered and normalised: entities, dates, figures, units, relations and temporal context. Then it is decided which part of that evidence matters for this commission, this audience and this moment. Only then does the narrative structure appear, and generation writes over it.
The change has three effects we have been able to observe. A correction to the data propagates on its own to the text version, the narrated one and the subtitles, because all three are produced from the same approved script. The automatic checks stop being language analysis and become comparisons against the source structure. And human review concentrates on what really does require judgement: what has been left out, and whether the narrative order supports what the evidence allows you to claim.
It is not a faster method for the first piece. It starts to pay off with the second, and it clearly pays off when the same information has to exist in several formats and languages.