# Palimpsest

> Palimpsest is an open-source, public-good observatory that measures how
> authoritarian states — and now their state-aligned AI models — censor. It
> treats deletion as data: it archives public posts, watches for their scrubbing,
> and ranks what the censor is most focused on right now. It also measures the
> "generative firewall": the share of sensitive answers that state-aligned AI
> models refuse or rewrite, proven selective against neutral controls.

Palimpsest watches the censor, never the censored. It is free, MIT-licensed, and
never monetizes the people or topics it observes. Every finding ships its raw
evidence, so journalists and researchers can verify each claim.

## What Palimpsest measures

- **Deletion as data.** Which topics the Chinese censor is most actively scrubbing
  right now (censorship velocity), validated against six documented censorship
  events.
- **The Generative Firewall Index.** What a state's AI is engineered to hide — the
  refusal / narrative-substitution rate of state-aligned large language models on
  sensitive prompts, in Chinese and English, against Western control models.
- **The GDELT cross-signal.** Whether the topics being deleted inside China are
  simultaneously loud in global news coverage: censored at home, loud abroad.
- **GitHub-as-Refuge.** Takedown pressure on GitHub mirrors of censored material
  (996.ICU, nCovMemory and others), judged only against persisted prior-presence
  baselines so a takedown claim is never fabricated.
- **Safety as architecture.** Public reads only; no person inside is ever asked to
  act; no state-aligned model is ever the analyst.

- **Wayback Reconstruction.** Deletions and silent redactions of public Chinese URLs,
  recovered from the Internet Archive's capture timeline. The Archive is a retroactive,
  outside-the-wall observer: for any URL it crawled, its CDX index brackets a takedown or a
  state rewrite with real, archive-witnessed timestamps and a permanent citable snapshot on
  each side — recovering deletion velocity from open egress, reported fail-loud as an explicit
  capture bracket.

All five signals are live and self-updating on schedules from 3 hours to daily, via
public GitHub Actions on the open repository.

## Key pages

- [Observatory](https://palimpsest.info/dashboards/ddti_observatory.html): live fear index, topic network, provenance-on-pull
- [Generative Firewall Index](https://palimpsest.info/readings/generative-firewall-index.html): live index of state-aligned AI suppression
- [Raw evidence (JSON)](https://palimpsest.info/readings/latest.json): machine-readable latest readings
- [For researchers](https://palimpsest.info/for-researchers.html): every live feed, its schema, honest scope, and how to cite (BibTeX)
- [All readings](https://palimpsest.info/readings/)
- [Source code & method (GitHub)](https://github.com/beepboop2025/palimpsest)

## Citation

Cite as: Palimpsest — an open observatory of authoritarian censorship,
palimpsest.info, accessed [date]. Open-source (MIT). Every finding ships raw evidence.

## Frequently asked

**What is Palimpsest?** A free, open-source (MIT) public-good observatory that measures how authoritarian states censor by treating deletion as data, and measures how state-aligned AI models refuse or rewrite sensitive answers. It watches the censor, never the censored.

**What is China censoring right now?** Palimpsest archives public Chinese posts, detects deletions, and ranks the topics being scrubbed fastest as a live censorship index at palimpsest.info, with raw JSON evidence.

**What is the Generative Firewall Index?** A live measure of how often state-aligned large language models refuse or substitute state-narrative answers on sensitive prompts, in Chinese and English, against neutral control models — proving suppression is deliberate and selective.

**Is Palimpsest free and open source?** Yes. MIT-licensed, fully open source, never commercial, developed in the open at github.com/beepboop2025/palimpsest.