Open-source · public good · MIT
Palimpsest
An open observatory that measures how authoritarian states censor, and a sealed public registry of AI evaluation results. Deletion is treated as data. Every result seals the moment it goes up, hash-chained, so nobody can revise it afterwards. Not a lab, not a government, not us. It watches the censor, never the censored, and every number traces to public evidence.
Verifiable Eval Registry
AI evaluations nobody can revise after publication. Questions frozen before any model is asked, every result hash-chained. Chinese state-aligned and Western frontier models on the same record, side by side, watched for refusal drift over time.
Open the registry →Erasure Observatory
What the record lost, measured across three layers: network, narrative, and model. Each reading seals into a tamper-evident ledger, so our own history cannot be quietly rewritten either.
See the ledger →DDTI Observatory
The full command surface: the censorship fear index, a live topic network, ranked censor targets, and the GDELT and GitHub-refuge signals in one view.
Open observatory →Generative Firewall Index
What a state's AI is engineered to hide: the share of sensitive answers state-aligned models refuse or rewrite, against neutral controls, in Chinese and English.
View the index →Censored at home, loud abroad
The GDELT cross-signal: which censored topics the world's press covers heavily while they are being scrubbed at home. Containment, made visible.
See the signal →Pressure on the mirrors
GitHub-refuge watches the repos that shelter censored material, 996.ICU to nCovMemory, for takedowns, legal blocks, and defensive preservation bursts.
See the watch →For researchers & journalists: data & methods →
What it does
- Deletion as data. Archives public posts, watches for their scrubbing, and ranks what the censor is most focused on right now. Validated against six documented censorship events.
- The generation layer. Probes state-aligned LLMs and records their refusals and state-narrative substitutions, in Chinese and English, against Western controls.
- The sealed registry. Every evaluation is pre-registered and hash-chained the moment it is published. A lab, a government, or Palimpsest itself cannot revise a number afterwards, and anyone can recompute the chain to check. Western frontier models now sit on the same record as the state-aligned ones, watched for what they quietly stop answering.
- Safety as architecture. Public reads only, no person inside is ever asked to act, and no state-aligned model is ever the analyst. Every finding ships its raw evidence.
Explore
For researchers & journalists: data & methods → Source code & method (GitHub) How it is measured Security & safety All readings Support the observatory
Frequently asked
What is Palimpsest?
Palimpsest is a free, open-source public-good observatory that measures how authoritarian states censor by treating 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: how often state-aligned AI models refuse or rewrite sensitive answers. It watches the censor, never the censored, and ships raw evidence with every finding.
What is the Palimpsest censorship observatory measuring?
Two things. First, deletion as data: which topics the Chinese censor is most actively scrubbing right now, measured as censorship velocity and validated against six documented censorship events. Second, the Generative Firewall Index: the share of sensitive answers that state-aligned large language models refuse or rewrite, tested in Chinese and English against neutral Western control models to prove the suppression is deliberate and selective.
What is China censoring right now?
Palimpsest tracks this live. It archives public Chinese posts, detects when they are deleted, and ranks the topics being scrubbed fastest as a real-time censorship index. The live readings and raw JSON evidence are published here, updated continuously.
What is the Generative Firewall Index?
The Generative Firewall Index is a live measure of what a state's AI is engineered to hide. It records how often state-aligned large language models refuse or substitute state-narrative answers on sensitive prompts, in Chinese and English, against neutral control models. A high index means the model is selectively suppressing truthful answers by design, not by accident.
What is the Verifiable Eval Registry?
The Verifiable Eval Registry is a public, tamper-evident record of AI model evaluations. The questions are frozen and sealed before any model is queried, each result is hash-chained to the one before it, and the whole registry is fingerprinted by a Merkle root. A result whose questions were not frozen first, or a number edited after publication, fails verification. It currently holds sealed audits of Chinese state-aligned models and Western frontier models side by side, on the same frozen questions, and anyone can recompute the chain from the public repository.
Is Palimpsest free and open source?
Yes. Palimpsest is MIT-licensed, fully open source, and developed in the open as a public good. It is never a commercial product and never monetizes the people or topics it observes. Its code, method, and raw evidence are all public on GitHub.
Free and open-source, developed in the open as a public good. It is never a commercial product and never monetizes the people or topics it observes. The internet stays free because people keep measuring the dark.