A story, not an event

Headlines cannot separate “Starmer will resign” from “Starmer has resigned” from “the race to replace him” — measured at 0.86 precision for stories against 0.23 for events, and human annotators do no better.

Grouping runs in two passes: a broad one that finds themes, then a tighter one inside each theme that separates the stories.

Sources counts voices, not articles

One per distinct publisher, after collapsing reprints that share a headline word for word. Between a fifth and a third of clustered articles are wire copy — six local papers running one agency line are one voice, and counting six would make disagreement look smaller than it is.

Polities are the unit of comparison

Outlets are grouped by the polity they publish from, because that is a fact rather than a judgement about editorial stance. Filled marks covered the story; hollow ones stayed silent.

A story needs two polities to be measurable — one voice cannot disagree. Real outlets publish in bursts, so silence often means “not this hour”, not “not working”.

The fixed panel is not chosen yet, so the number here counts the polities that actually carried the story, and the globe plots them from the data. Hollow marks are candidate polities that stayed silent — illustrative until the panel exists.

Named by, and the omission

How many sources put that actor in the headline. The signal is usually what is missing: an outlet reporting deaths without naming who caused them has made an editorial choice, and it is legible without anyone judging tone.

Three states, not two: present, absent, and not evaluable — the last when no alias exists in that script. Open a story to see every source and check the count yourself.