Reading real clustered output caught a config error the tests could not:
Ohio Capital Journal and Kentucky Lantern are both States Newsroom outlets
syndicating the same national desk, and carried 10 byte-identical headlines in a
single poll -- more than any other source pair. Configured as separate buckets
they would have doubled the corroboration signal on every syndicated story,
which is exactly the failure the bucket concept exists to prevent.
Found by checking identical headlines across all source pairs rather than
spot-fixing the one noticed by eye. That sweep also confirmed the buckets
already set deliberately are right: BBC Top/World, Ars Technica/Science and
NPR News/Politics each share a bucket and each showed shared headlines.
Regression test added.
connor
d047c58d8a