# Outliers

Every 311 complaint filed in New York City, read by Jev (TypeSafe's System One decision model)
as it lands. Jev answers 5 fixed questions per complaint and returns probabilities, not prose.
A deterministic formula turns those answers into a weirdness score from 0 to 100. Complaints
scoring 40 or more are posted. Nobody votes; the ranking is the model's alone.

## Questions asked of Jev (one call per complaint)
- surprise  Score  mundane / routine / notable / strange / bizarre
- risk      Score  none / minor / moderate / serious / severe hazard to health or safety
- place     Noul   is this complaint type unexpected for this location type
- pair      Score  ordinary .. unheard-of pairing of descriptor and complaint type
- category  Choice animals / water_waste / noise / vehicles / wildcard / none

## Score
weird = rare + 12.5*surprise + 1.25*risk + 10*place + 2.5*pair  (parts sum to 100, no cap)
rare is 0..25 from the complaint type's citywide count over the trailing 30 days (log scale).

## Second feed: Same story, different words (/news)
Headlines from about 30 outlets' RSS feeds, grouped into stories by TF-IDF cosine similarity over
a 48-hour window (not by Jev). Jev reads each outlet's headline and teaser (its RSS summary) with the outlet name
withheld and answers 5 fixed questions:
- tone       Score  reassuring / neutral / concerned / alarmed / sensational (wording, not the event)
- claim      Choice reports_fact / attributes / asks / opinion
- blame      Choice none / government / company / individual / group / foreign_state / nature
- overreach  Noul   the headline claims more than the summary supports
- loaded     Noul   emotionally loaded or dramatic wording
divergence = 25*tone_spread/4 + 25*blame_split + 25*overreach_spread + 25*claim_split, for stories
with 3 or more scored outlets. Stories over 30 post. JSON at /api/news.json.

## Data
- Feed JSON: /api/scores.json (top posts, last-24h totals, formula)
- Source: https://data.cityofnewyork.us/resource/erm2-nwe9.json
- The city publishes with about a one-day lag; the feed replays each day at its own pace, 24h behind, with true timestamps.
- Addresses are withheld. Text is the city's own picklist values.

Belvedere Labs, a Clear Fracture project. Runs as a Belvedere pipeline.