Same story or different event? How the pipeline keeps sources apart

Two fines, one company

In March 2025 two stories about the same regulator ran a week apart: a fine for a data-transfer breach, and a fine, against the same company, for cookie consent. Same regulator, same company, same month, different amounts. A draft that treats those as one story averages the amounts, or picks one and attributes it to everybody.

The pipeline's second step exists for this. Before anything is read in full, search results are grouped by the event they describe, using the names, dates and figures in each snippet. Results that share an origin, a wire paragraph carried by four papers, collapse into one. Results that share a subject but not an event stay apart.

What grouping changes

Grouping happens before reading because reading is what costs money and time. Fifty search results become six page fetches: one per origin, in the group the brief is about. The other group is kept, and named in the run trace, so the editor can see that a second event was found and set aside.

It also changes the claim check. A claim is corroborated when two independent origins in the same group support it. An origin from the other group cannot corroborate it, even if the sentence matches, because it is describing a different thing.

When it goes wrong

Grouping fails in two directions. Merge too eagerly and two events become one; split too eagerly and a real corroboration is thrown away. The pipeline errs toward splitting, because a missed corroboration produces a flag a person can resolve, while a false merge produces a confident sentence that is wrong.

Frequently asked questions