EXPERIMENTS · BUILT 2 Oct 2026, 00:35 UTC

Which stories are the same event?

An edition can carry several stories about one thing that happened. How well does the rule that groups them work?

Question: Which of an edition's stories are the same real-world event?
Method: Centroid similarity, shared principals, Google's grouping and time apart, per pair; star grouping around the strongest story. Scored against pairs labelled same or different by hand.
Result: Precision 0.92, recall 0.338 on 265 labelled pairs. Latest edition: 105 stories shown as 91 events.
Decision: Keep the rule: it is precise, and no threshold in the sweep raises recall by more than a few points while keeping precision above 0.9. The pairs it misses are different angles on one event, which similarity alone does not see; catching them needs another signal.

How it was measured

An edition can carry several stories about one thing that happened: a summit's arrival and its trade outcome, say. The site shows such stories once, the strongest keeping its place and the rest listed beneath it as related developments. That changes only how an edition is shown: no story, rank or score changes.

For each pair of stories in an edition, the rule looks at how similar their articles are, which principal people and organisations both name, whether Google News grouped them, and how far apart they ran. Two stories are one event at similarity 0.9 alone, 0.82 with 3 shared names, or 0.7 when they name the same principals with weight. Groups form around the strongest story, so a chain of loosely similar stories never merges.

The rule is scored against pairs labelled by hand as the same event (including a later development or a different angle on it) or different (related, but separate things that happened). Precision is how often the rule is right when it groups two stories; recall is how many same-event pairs it finds. The target is precision above 0.9, because a wrong merge hides a separate story.

The numbers

0.92precision: right when it groups

0.338recall: share of same-event pairs it finds

265labelled pairs, 136 of them the same event

Latest edition (2026-10-01): 105 stories shown as 91 events.

Other thresholds · the same labels, the rule's three similarity thresholds moved

Alone	With names	Same principals	Precision	Recall
0.86	0.78	0.66	0.841	0.39
0.86	0.78	0.7	0.855	0.39
0.86	0.78	0.74	0.85	0.375
0.86	0.82	0.66	0.891	0.36
0.86	0.82	0.7	0.907	0.36
0.86	0.82	0.74	0.904	0.346
0.86	0.86	0.66	0.907	0.287
0.86	0.86	0.7	0.929	0.287
0.86	0.86	0.74	0.925	0.272
0.9	0.78	0.66	0.847	0.368
0.9	0.78	0.7	0.862	0.368
0.9	0.78	0.74	0.857	0.353
0.9	0.82	0.66	0.902	0.338
0.9	0.82	0.7	0.92	0.338
0.9	0.82	0.74	0.917	0.324
0.9	0.86	0.66	0.923	0.265
0.9	0.86	0.7	0.947	0.265
0.9	0.86	0.74	0.944	0.25
0.94	0.78	0.66	0.83	0.324
0.94	0.78	0.7	0.846	0.324
0.94	0.78	0.74	0.84	0.309
0.94	0.82	0.66	0.889	0.294
0.94	0.82	0.7	0.909	0.294
0.94	0.82	0.74	0.905	0.279
0.94	0.86	0.66	0.909	0.221
0.94	0.86	0.7	0.938	0.221
0.94	0.86	0.74	0.933	0.206

Who labelled, and limits

255 of the 265 pairs were labelled by Claude, an AI model, and 10 by the site's author. Mostly one labeller's judgement: another reader would draw some lines differently, particularly on what counts as a different angle on one event.
The pairs come from a few days of editions and lean toward hard cases: stories that name the same people, or that Google News split.
A pair is labelled for the day it ran: the same two stories can change overnight as new articles arrive.