# ForecastGeo

> News stories placed on a map by where they happened, plus four things computed on top
> of that archive: side-by-side comparisons of how different outlets framed the same
> story, coverage geography, a scored forecast ledger, and per-outlet statistics.

ForecastGeo ingests 2,690 articles from 51 newsrooms and
extracts the location of each story. The archive begins 30 July 2026 and is retained
permanently. Article pages restate publisher headlines and summaries and link out to the
original; the pages listed under "Derived analysis" are this site's own work and are not
available elsewhere.

## Derived analysis

- [Framing comparisons](https://forecastgeo.com/coverage): 81 stories currently covered by two or
  more outlets, each with a per-outlet breakdown of what that report leads with, whose
  account it credits, and which shared claims it leaves out. Individual pages are at
  `/cluster/:slug`.
- [Coverage geography](https://forecastgeo.com/coverage): attention share across 55 countries,
  an outlet-by-country matrix, how long each country has gone unmentioned, and which
  countries appear only in bursts.
- [Forecast ledger](https://forecastgeo.com/ledger): scenarios generated per story with a resolution
  deadline, checked against later coverage, scored with Brier scores and a calibration
  table. Unresolved runs are reported as unresolved and never counted as hits.
- [Outlets](https://forecastgeo.com/sources): per-newsroom pages — where each reports from, whether its
  article text is readable, how often it credits sourcing, and its image-audit record.
  Individual pages are at `/source/:slug`.
- [Image audit](https://forecastgeo.com/images): which article photos a vision model judges generic or
  unrelated to their story, and which identical files ran at more than one outlet.

## Data

- [JSON API](https://forecastgeo.com/api): every page above is also readable as JSON. No key required.
- [Atom feed](https://forecastgeo.com/feed.xml): newest stories.
- [Sitemap](https://forecastgeo.com/sitemap.xml)

## How to read the numbers

- Model-written text is labelled on the page with the model that produced it. It is
  confined to restating, comparing or classifying supplied articles — never to supplying
  facts of its own — and every user-visible claim cites the article it came from.
- "The same image" means two files with an identical SHA-256. No perceptual matching is
  used, so this undercounts re-encoded wire photos rather than guessing.
- Coverage figures describe the retained window, which begins 30 July 2026. They are
  not long-run trends and the pages say so.
- Roughly half the archive has no readable article body: some publishers decline
  crawling in robots.txt, and that is recorded as declining rather than as an error.