Transparency

How the observatory works, what it keeps where, who may use what, and where every method and its limits are written down. Three layers, each with its own access rule.

  1. BronzeOriginals as collected: every Telegram message object with forwards, edits, deletions and media, every web page as fetched, with provenance and earlier versions. Not downloadable; on request.
  2. SilverCleaned and indexed data: deduplicated, language-detected, transcribed, placed, labelled, counted. Published as versioned files with a dictionary, checksums, citation and terms under Data.
  3. GoldResearch after review on this site: the incident studies and the documents under Publications, each with its status, method and the data it rests on. Not a synonym for peer review.

The pipeline

Seven steps, each explained in plain language with what it adds and how reliable it is: collect, clean, near-duplicates, labels, places, source graph, weekly counts — on Coverage and methods. Whether each stage ran recently is on Public status.

Access model

Validation and reliability

Three reliability measurements are reported and never merged: independent round-1 inter-coder reliability, post-adjudication residual agreement and machine-versus-gold accuracy. The coding guide is the rule book behind them. Where a step is not yet reliable the site says so and does not count on it.

Method documents

Limitations and corrections

Each dataset page lists its limitations; each study carries its cannot-claim list; the roadmap records the project's phases and gates as they stand. Every change to displayed data is logged with its run ids in the changes log; corrections are logged the same way and never silently.

Public status · About · Terms of use