Connor spotted a Missouri redistricting story labelled ohio-state-politics.
Measuring showed that was not an isolated slip: 66% of the winners of a
place-based block named no place at all. Glastonbury, Anthony Albanese and a
Tim Montgomerie story were all scoring as italy-mexico; a singing-apraxia story
was ohio-state-politics.
Embeddings model meaning well and named entities poorly -- "Italian politics"
and "Australian politics" are nearly the same vector. Place is identity, not
meaning, so it is matched lexically, which is the same division of labour that
makes lexical comparison right for syndication detection and wrong for
clustering. Blocks carry `requires_any`; a block cannot win for an item whose
text names none of its terms.
Word boundaries, not substrings: "rome" is inside "chrome" and "mali" inside
"normalise". That is the trap the old blocklist fell into with "nfl" inside
"conflict".
The first implementation was wrong in a way worth recording, because it looked
correct and measured almost unchanged (66% -> 30%). It wrote a sentinel into the
RAW score before normalisation -- but when most items are gated, that sentinel
becomes the block's own mean, so a gated item z-scored to roughly zero and could
still win. Normalisation erased the gate. Applied after normalisation instead:
66% -> 0%, and place-block winners fell from 140 to 10, so 130 items now land in
the block they actually belong to.
