What is Palimpsest for?

Palimpsest helps researchers, journalists and institutions inspect China information-control evidence across attributed reports, public social context, deletion records and defined network measurements, with freshness, limits and revisions attached.

Is Palimpsest a newspaper?

No. Publisher reports remain attributed source reports. Palimpsest adds measurement context, source grouping, revision records and explicit limits; it does not convert a report into independently verified fact.

Why did the AI evaluation work begin?

Palimpsest's founder began with a screenshot of a state-aligned model changing or withholding an answer about criticism of the Chinese Communist Party. That observation became a frozen, repeatable evaluation because one screenshot does not prove a pattern and does not prove motive.

China information-control evidence Open source · no paywall

Palimpsest × NarcoScope research desk

Follow China’s economy.Examine the public record.

Explore China’s industry and housing data, changes in official publications, and the regional questions around CPEC, Balochistan and the Belt and Road. Open NarcoScope’s global drugs, arms and informal-economy records in the same research workspace.

Open China evidence Explore world data Open the weekly evidence brief

Read the weekly brief or register your research interest in Telegram. Starting the bot also subscribes you to daily updates; /stop ends them. A click is not counted as a subscriber.

Not a newspaper. A publisher report remains a publisher report. Palimpsest only labels what its own evidence adds.

Reproduce. Clone the repo and run python3 scripts/reproduce_all.py. Exit 0 means the sealed chains and derived scorecards still match.

Current acquisition route

Evidence roles stay separate

China situation desk

One view. Three evidence layers. Limits intact.

Open attributed reporting, reviewed public circulation context and Palimpsest measurements in the same dossier. Context never silently becomes corroboration.

3 declared evidence roles

Open: Read the current deskpublisher · social · measurement
Limit: Topic overlap is context, not article verification. Missing evidence remains a visible gap, never a zero.
Method: Deterministic synthesis from attributed reports and checked public records; no generative summary is published as evidence.

Start with a research question

Three ways into the evidence.

China economic health

Go beneath national headlines

Compare industries, city housing and source tables. Keep each observation’s reference period and definition visible.

Open the economic desk → Shared data library

Compare countries and histories

Explore drugs, arms and informal-economy indicators with maps, time series and downloadable source records.

Explore the data → CPEC, Balochistan & BRI

Follow the regional questions

Examine national finance, project scrutiny and local reporting. Shared geography keeps the evidence connected without establishing a causal claim.

Open regional research →

Continue from the China desk

Inspect, follow or reproduce.

The primary route stays focused. Adjacent tools remain available when the evidence task requires them.

ReadThe current China situation deskOpen evidence view → TraceA deleted post or changed public recordOpen evidence trails → MeasureDefined China-facing public signalsOpen observatory → CheckA suspicious Telegram messageOpen checker → VerifyA sealed AI evaluation runOpen registry → MapBelt and Road projects, economics and local impactsOpen coverage ledger → BuildWith the read-only API or MCP serverOpen developer tools → FollowOnly the feed you actually needChoose RSS or JSON → CiteA specific signal on a specific dayOpen citation pack →

Live publication state

Current results, not product claims.

Each row separates what the system currently reports from the file that supports it and the sentence that keeps it bounded.

AI eval registry

Sealed runs: checking current receipt

Receipt

Verify the chainreceipt unavailable

Limit

A valid chain makes later edits detectable. It does not prove that an evaluation construct is valid.

China observatory

Signals: checking/checking

Receipt

Inspect every signalchecking current receipt

Limit

These are measurements of defined public signals. Movement is not proof of intent or a single cause.

Source index

Reports: checking current receipt

Receipt

Open the labeled indexSource coverage: checking/checking

Limit

A publisher report is not a Palimpsest finding. Single-source entries are labeled and link back to the original.

Product boundary

What this is for.

If you only need a newspaper article, read the publisher. Use Palimpsest when you need a measurement, a source relationship, a revision trail, a machine-readable record, or a clear statement of what remains unknown.

PublishesOwn measurements, scoped analysis, receipts and public data

IndexesAttributed publisher reports without relabeling them as verified facts

Does notReplace original reporting, infer motive from co-movement, or hide failed inputs

Syndication

Every feed has one job.

The directory states who each feed is for, what it adds, and what it cannot establish before you subscribe.

Live eval resultsNew sealed-run findings and revisionsRSS Eval methods journalHuman-readable method changes and failuresRSS Measurement updatesPalimpsest's own current instrument resultsRSS Source index + measurementsClearly labeled publisher reports and instrument updatesRSS

Compare all RSS and JSON feeds → Read the eval methods journal →

External evidence relay · Dated fallback

What the 2025 INCB precursor report says about China-linked incidents: and what it does not prove

Nine reported incidents, nearly five tonnes, and one crucial limit: reported origin is not causal attribution.

Data as of: 12 Aug 2026
Incidents: 9
Upper-bound mass: nearly 5 t
Visual citation coverage: 100%

Read the NarcoScope investigation · Open NarcoScope

Evidence network

Useful together. Never collapsed together.

These are direct routes to separate public tools. A relationship between records is not a merged claim.

Palimpsest

Public observatory readings and citations.

Daily bot · Signal desk

NarcoScope

Official-record evidence on illicit-market infrastructure.

Website · Evidence bot

ScamShield reviewed bridge

scamshield-intelligence-pack/v1, release 2026-08-08.2, maps 18 public sources to 8 typologies. HUMAN_REVIEW_REQUIRED is mandatory. The levels describe evidence relationships, not guilt. Raw Telegram text is hashed and privacy-minimized candidates are never auto-published.

Public pack · ScamShield bot

AI assistance, named

Automation is useful when its responsibility is visible.

AI assistance accelerated this rebuild and supports parts of the development workflow. It is not presented as evidence. Published surfaces distinguish original sources, deterministic analysis, machine-authored text and human review; the code and tests are public.

Palimpsest's founder began the evaluation work after a screenshot showed a state-aligned model changing or withholding an answer about criticism of the Chinese Communist Party. One screenshot could not show a repeatable pattern and does not prove motive, so the observation became a frozen evaluation and a revision-safe registry. Inspect the machine-readable assurance report.

See what changed after feedback Inspect the source