Palimpsest · data directory
Every dataset the observatory publishes, at stable URLs. Each reading ships its raw evidence beside the headline number, so any figure on this site can be audited from the files below. Schemas, honest scope, and citation guidance: for researchers.
| Signal | Cadence | Latest | Time series | View |
|---|---|---|---|---|
| DDTIdomestic discourse tightening index — deletion signals | every 3h | ddti-latest.json | ddti-history.jsonl | observatory |
| Generative Firewall Indexrefusal / state-narrative tomography of state-aligned LLMs · k-sampled with 95% Wilson bands | daily | latest.json | history.jsonl | dashboard |
| GDELT cross-signalcensored at home, loud abroad — containment detection | every 6h | gdelt-latest.json | gdelt-history.jsonl | observatory |
| GitHub-as-Refugedeveloper migration toward censorship-resistant infrastructure | every 12h | github-refuge-latest.json | github-refuge-history.jsonl · baselines | home |
| Date | Reading | Artifacts |
|---|---|---|
| 2026-07-01 | Generative Firewall Indexrefusal / state-narrative tomography of DeepSeek & Qwen vs a Western control, sensitive Chinese-language probes plus neutral controls | reading · dataset |
How a reading is made. A human-ratified probe set (never model-derived) is put to a model panel through a public API; every response is recorded verbatim; all judgement is the repo's lexical, auditable rule set — no state-aligned model is ever the analyst; the Chinese models are the subjects under observation. Public reads only; no jailbreak. When a source breaks or a number cannot be stood behind, the signal is suppressed or the run is skipped — never faked. Machine-readable summary for AI agents: llms.txt. Complete file tree including newer dated readings: github.com/beepboop2025/palimpsest/readings. All data is free to reuse with attribution (MIT).