Connor noticed 5 of 5 alerts were Middle East. Measuring showed that was not
proportionate: Middle East blocks are ~11% of the eligible corpus while
software-systems alone is 25%, yet Middle East took 35% of the top 20 and
software took none.
Two mechanisms, both measured:
1. Five conflict-* blocks could each claim their own alert slot. Interest blocks
now carry a `group`, and routing enforces max_alerts_per_group_per_day, so
related blocks share one quota. This alone moved the mix from 5/5 Middle East
to three distinct groups.
2. Z-scoring subtracts each block's MEAN, which silently encodes how much of the
corpus that block matches. software-systems has the HIGHEST mean raw
similarity of any block (0.537) and the LOWEST max z (2.51);
conflict-mideast-israel has a lower mean (0.460) and reaches 3.47. A topic is
penalised for being well covered in your own feed list.
Added `normalize = "top-anchored"` as an alternative that rescales each block by
its own upper tail instead of its mean, leaving `weight` to express preference
rather than corpus composition.
Also ruled out a plausible-sounding cause rather than assuming it: HN and
Lobsters items are title-only and much shorter than news articles, so terse
titles seemed likely to match prose probes poorly. Measured the opposite --
title-only HN items score 0.618 against the software block, ABOVE Phoronix/LWN
with summaries (0.610) and Ars Technica (0.540). Short text is not the problem;
prevalence is.
Scoring (news/score.py): every profile block is embedded and compared by
z-score WITHIN block, then weighted. Veto fires on absolute similarity OR a
margin over the interest signal; demotes apply a penalty instead of excluding,
because those categories sit adjacent to real interests.
Importance is deliberately NOT corroboration count -- the old system's error,
which measured the subscription list rather than the news. It combines block
similarity, corroboration, burst velocity, editorial prominence and aggregator
score, all weighted from config.
Routing (news/route.py): two lanes with separate budgets. Interest never
overrides quiet hours; importance may, within a hard nightly cap. Vetoes are
absolute on BOTH lanes -- the v0.2 bypass lane never consulted relevance, so a
blocklisted item could still fire at priority 5, and a test now asserts it
cannot. Budgets are spent strongest-first rather than first-come.
Measured on 1,318 real clusters: 2.5 notifications/day and a 14-item digest,
inside the 2-5/day target.
Three defects the real data exposed, none of which unit tests would have caught:
- velocity was computed as origins/hours with a 0.5h floor, giving every
SINGLE-source cluster a spurious velocity of 2.0. Growth rate is undefined
for a cluster that has not grown; it is now zero below two origins.
- deferred items (quiet hours, budget exhausted, immature) settled straight
into the digest and bypassed its cap, producing 198 items/day. Everything
that misses an alert lane now competes for one ranked, capped digest.
- feeds return large archives on first fetch -- 88 Strong Towns items going back
five months, and a BBC entry titled "BBC News app" dated 2025. Those belong in
the scoring population but must never notify, so routing now has a recency
requirement.
Reading the output also found two profile gaps: recurring columns and link
round-ups ("Catching Our Eye News Roundup", "Astronomy Picture of the Day")
scored as top-ranked interest items. They are on-topic but publish on a
schedule, so they are demoted rather than vetoed and still reach the digest.
Connor's point: "several outlets carried this" covers situations that carry
different amounts of information, and origin_count() flattened them all to one
number. Splitting on same/different newsroom and same/different text gives four
cases:
same newsroom, same text -> pure duplicate (BBC World + BBC top). Zero
signal; the front-page placement it does carry is
already captured by sources.prominence.
same newsroom, diff text -> DEPTH. A newsroom filing four pieces on one story
is a real editorial judgement, previously
discarded entirely.
diff newsroom, same text -> PICKUPS. Syndicated copy; each carrier chose to
run it. Weak but non-zero.
diff newsroom, diff text -> ORIGINS. Independent judgement and independent
newsgathering. Strongest.
score = origins + 0.3*log1p(pickups) + 0.2*log1p(depth), weights in config.
Origins stay dominant; the others break ties among the many two- and
three-origin clusters, and log damping stops a high-volume feed inflating depth.
Measuring first changed the design. Depth outnumbers pickups 78 to 11 across 91
real clusters, because these sources are mostly national outlets writing their
own copy rather than local papers running wire -- so the more valuable half was
the one not originally proposed. UK settlements scores 10.55 from 10 origins
plus 15 same-newsroom pieces; AfD scores 5.55 from 5 origins plus 15.
Tests assert the four-case ordering directly, and that twenty pieces from one
newsroom still rank below two independent ones. Both needed a helper generating
vectors with exact pairwise cosine: the band between the 0.82 clustering
threshold and the 0.97 syndication threshold is narrow, and two earlier drafts
of these tests used ad-hoc vectors that fell outside it and asserted nothing.
Connor's observation: two outlets running the same story can be caught by
comparing the text, without anyone having to know who owns whom.
Corroboration is now origin_count(), a union-find where two items share an
origin if EITHER they are in the same bucket, OR their text is near-identical
(normalised title match, or cosine >= 0.97). The rules are complementary:
buckets are same-owner/different-text (NYT filing five distinct AfD pieces),
text matching is different-owner/same-text (a local station reprinting wire).
On real data it fires exactly where it should: the vote-by-mail cluster
collapses 2 buckets -> 1 origin, catching the States Newsroom syndication
automatically -- the case previously fixed by hand in config -- and a
Smithsonian story collapses 3 -> 2, catching WCPO reprinting wire, which no
static bucket can express since WCPO is genuinely independent on local stories.
The eight Renoir-heist reports, each independently written, correctly stay at 8.
Across 91 multi-item clusters: 174 buckets -> 163 true origins.
Note this is the one place a lexical comparison is right. Clustering fails on
lexical methods because paraphrases share no wording, which is why the SimHash
stage was measured and deleted -- but syndicated copy is literally the same
bytes.
First formulation was wrong and the data caught it: summing syndication groups
and solo buckets produced 3 origins for a 2-bucket cluster. Syndication may only
ever reduce corroboration, which the union-find guarantees and a test asserts.
Audit found three things compiled in that shouldn't have been: the User-Agent
carried the author's domain, the cluster threshold and window were module
constants the pipeline never overrode, and the embedding model was pinned --
which blocked the one change most likely to be wanted, swapping in a
multilingual model for native-language sources.
Configuration is now three files with distinct roles: config.toml (machinery),
profile.toml (who the reader is), sources.toml (what to read). All are
relocatable via NEWS_CONFIG_DIR. news/*.py now contains no interest, place name,
feed URL, or personal identifier.
tests/test_config.py keeps it that way: it builds a configuration for an
invented reader and asserts none of the author's identity or preferences reaches
the running code. Writing it immediately found a real bug -- CONFIG_DIR was
resolved at import time, so NEWS_CONFIG_DIR silently did nothing if set after
the module was first imported. Now resolved per instantiation.
Documented the coupling that portability creates: [clustering].threshold and
[scoring].veto_absolute are cosine constants calibrated against one corpus with
one model, so changing [embedding].model invalidates both.
connor