More public vantages, continuously, then join them.

The join

Sit on warehouses incrementally. Then join what you already sealed.

These rows are projections of readings Palimpsest already published. They are not a new scrape. A shared host is not an exact URL join. Official text alone is the censor's story.

Host-surface joins

Wire headline, same host blocked in OONI.

This is the closest sealed pair today. It is still not the four-leg tuple.

Captured demand

Query demand, not posts. Titles Palimpsest already captured.

Archive watch

Still in Wayback, or not yet witnessed.

OONI host anomalies

Blocked in OONI, at host grain.

Four methods that change the volume

Drink firehoses. Watch a URL list. Sit on warehouses. Never keep the HTML.

Drink firehoses. A public stream pushes every event. You filter with the gazetteer and write one line. That is how you get bulk without hitting a thousand sites. Wikimedia EventStreams, Bluesky Jetstream, Certificate Transparency.

Watch a URL list the cheap way. Keep a reviewed set of public URLs. Every few seconds: HEAD or a tiny GET, store status and a hash. When the hash flips, that is the live event. Bodies stay out. This is how deletion research scales.

Take other people's warehouses incrementally. GDELT drops every 15 minutes. OONI, Censored Planet, IODA, and Common Crawl already do the heavy lift. Pull only the new slice. You already join some of these. You do not sit on them.

Never keep the HTML. Live bulk is NDJSON on the node: source, url, title, hash, time, vantage. Rotate daily. Python reads the tail. Git only gets a summary.

The legal grey stack

Sit on what the law forces out. Treat Wikipedia as a seismograph. Query demand, not posts. Read the country as numbers.

Listed-company announcements on 巨潮 are a public JSON firehose. Titles, times, PDF links. Same idea at city scale: 政府信息公开. Titles and links only.

Wikimedia EventStreams is a public SSE. Filter zhwiki plus gazetteer titles. You want deletes, revdels, and sudden protection, not every typo. Pair it with the hourly pageviews dump. You never asked anyone in-country.

Baidu, Toutiao, Douyin, Bilibili: titles and ranks only, abstain if a cookie wall appears. Do not chase signed search APIs or danmaku profiles. The signal is the rank flip.

OpenSky, OpenStreetMap minute diffs, AQICN and CNEMC-style station feeds. None of that is an article. Together it is a live physical pulse you can join to the wire when a city goes quiet, a plant vanishes off the map, or flights stop.

Chinese that lives outside the GFW: PTT, public HK boards, official CGTN/Xinhua YouTube titles, Bluesky Jetstream filtered on the gazetteer. Same Telegram rule: named surfaces, no login, never corroboration by itself.

Birth and death of names. Certificate Transparency for .cn and Chinese O-names. A new host today that is on a block list tomorrow is a live pair.

Coverage receipt

19 public vantages configured. 19 not attempted.

A zero is a coverage receipt, not silence. These doors write NDJSON on the node. Git only gets this summary.

What you do not add

Logged-in Weibo or WeChat is how people get hurt.

What you do not add: logged-in Weibo or WeChat, private messages, follower graphs, a browser session you did not own, or asking anyone inside to run a probe. That is not a stronger censorship tool. That is how people get hurt, and it is why CensorWatch is feature-flagged and isolated.