# Connor's news profile

The thing the previous system never had. `news_triage` v0.2's entire model of taste
was three keywords and two blocked substrings; everything else was ranking
machinery with nothing meaningful to rank. This file is the replacement, and
`profile.toml` is its machine-readable form.

Written 2026-09-08 from a structured interview plus evidence from 40 real feedback
clicks in the v0.2 database. Schema v2.

## How this is used

Every `description` and `probe` is embedded (bge-small-en-v1.5, CPU) into a vector.
An item's score for a block is the **max** cosine across that block's probes, never
the mean — averaging "Middle East geopolitics" with "grid-scale energy storage"
gives a vector near neither. Blocks are then compared by **z-score within block**,
because raw cosine is not comparable across blocks; see `CALIBRATION.md`.

Feedback moves these vectors over time. The written text is the cold-start seed, so
the system is useful on day one rather than after 200 clicks.

## Interests

**Foreign affairs is split in two.** Connor's own framing: *"my actual war interest
is low, but unfortunately all wars are important."* The click history looked
war-heavy only because world-news feeds are war-heavy. So statecraft —
`diplomacy-foreign-policy`, weight 0.9 — is the standing interest, while armed
conflict is weighted lower and routed mainly through importance.

Within conflict, his stated ordering became five weighted blocks:

| block | weight | |
|---|---|---|
| `conflict-anti-imperial` | 0.90 | anti-colonial and socialist-revolutionary wars |
| `conflict-us-involved` | 0.85 | wars the US is actively engaged in |
| `conflict-mideast-israel` | 0.80 | Middle East, especially engagements with Israel |
| `conflict-undercovered` | 0.75 | Sudan, DRC, Myanmar, the Sahel, Yemen |
| `conflict-general` | 0.45 | everything else — low interest, still important |

**Space and aerospace** (0.95) is the single highest-weighted block: Connor is an
aerospace engineer. He asked for launches and missions, deep technical coverage
(propulsion, materials, orbital mechanics, failure post-mortems), and
astronomy/planetary science — deliberately *not* the industry-and-contracts beat,
which his policy blocks already catch.

**Software and systems** (0.9, push-eligible) covers local model inference, Rust and
WebAssembly, self-hosting and infrastructure, and — in his words — *"Linux news in
general. The ricing world/neovim configs/etc."* Notably he did **not** pick security
and vulnerabilities, so that is absent rather than assumed.

**Political economy** spans `labor-inequality` (0.85), `left-movements` (0.85),
`anticolonial-sovereignty` (0.8), `far-right-watch` (0.75, tracked as a threat
rather than out of sympathy), and `political-theory` (0.7).

**Place**: `cincinnati` (0.9 — where he lives, so a lower bar), `ohio-state-politics`
(0.75), `kentucky-louisville` (0.6 — where he's from), `us-national` (0.8), and
`italy-mexico` (0.75).

Italy and Mexico are a special case. He follows both for politics but also because
he is learning the languages and expects to travel there, so **culture and popular
culture count as signal in this block, not noise** — the one place where lifestyle
content is wanted. The `consumer-lifestyle` veto must not swallow it.

Also carried: `science-discovery` (0.8), `environment-ecology` (0.75),
`tech-science-policy` (0.9), `urbanism` (0.8), `global-south` (0.7),
`energy-climate` (0.6).

## Vetoes, and why some things are demotes instead

Hard vetoes are reserved for categories that never overlap a real interest: US
big-three sports recaps, royal and celebrity gossip, consumer/lifestyle filler.

Anything that *might* overlap is a **demote** — a penalty on the score rather than
exclusion. This distinction exists because a hard veto on a sports-adjacent region
killed *"Belgian mayor blocks club from hosting Israeli opponent in Europa League"*,
which is Middle East politics. Overlap is the norm, not the exception:

- `crypto-ai-hype` sits next to AI regulation and local-inference news
- `culture-war-outrage` sits next to `far-right-watch`
- `true-crime-blotter` sits next to criminal-justice policy

### Explicitly not vetoed

Connor is not anti-sport; he is anti *US-big-three-wire-recap*. **Rock climbing,
chess, the Olympics and oddball sports are actively wanted.** Soccer is not blocked
as a concept — he is trying to get into it — though match previews and live threads
are noise. Sports *governance* (FIFA/UEFA corruption, Olympic host-city displacement,
Qatar migrant labour) reads as politics and stays.

## The importance lane

Independent of interest, and half the reason the project exists: war and escalation,
mass-casualty events, US federal shocks, economic breaks, and major US political
events (decisive elections, deaths of political figures).

**Vetoes still apply on this lane.** In v0.2 the bypass lane ignored relevance
entirely (`gate.rs:302`), so a blocklisted item could still fire at priority 5.

## Delivery

- Digest at **08:45**, quiet hours **22:00–07:00**. The importance lane **may** wake
  him; the interest lane never may.
- 2–5 pushes/day across both lanes.
- Conflict outside categories A–D notifies only on a major development — a new
  offensive, a ceasefire, mass casualties, a state entering. Routine front-line
  reporting goes to the digest.
- HN items are ignored below **150 points**: the crowd does the first filtering pass.
- The digest is a **complete record** of the day, including already-notified items,
  and the weekend edition runs longer and leans longform.
- **Three serendipity slots** in every digest for items that scored well on quality
  and importance but poorly on the profile, labelled "outside your usual". A system
  that only shows you what you already like tightens its own bubble as it learns.
- No format filter: he reads analysis, wire copy, investigations *and* opinion, which
  also confirms the old config's `notify_opinions = true` was deliberate.
- Feedback stays **binary**. The old system's failure was collecting 32 labels, not
  collecting shallow ones.

## Source trust

`weight` in `sources.toml` is editorial trust, carried over from the old config where
Connor had deliberately set Al Jazeera highest (1.2) and NYT lowest (0.8), and
extended to weight independent outlets (The Intercept, Drop Site, Tribune, African
Arguments, Middle East Eye, +972) at 1.1. It affects importance and tie-breaking,
never whether an item is visible.

## Language

Embedding is English-only. Measured: the same story in Spanish and English scored
0.758 and 0.600 cosine against a random-pair p99 of 0.634 — one of two translation
pairs was indistinguishable from an unrelated article. Native-language feeds would
rank erratically, so Italy and Mexico are covered by English-language sources.
Revisit with a multilingual model only if that becomes worth re-calibrating every
threshold for.

## Open questions

- Whether `global-south` is a genuine standing interest or an artifact of a small
  sample. Confirmed wanted by Connor, but worth re-checking against real feedback.
- Exact score at which the importance lane may override quiet hours.