Methodology
Plague Tracker breaks reporting into claims, which are single, checkable statements such as "Shelekhov district hospital placed under quarantine". Each claim links to the evidence behind it: a source (article, statement, post) and whether that source supports, contradicts or only gives context.
Two separate scores
Every claim gets a label that says how well sourced it is, and a percentage that is our best guess of how likely it is to be true. They can differ. A report from one very reliable news agency is likely true but still only "reported by one outlet". Two tabloids repeating a rumor make it "reported by several outlets" but the percentage stays low.
Labels
- Confirmed by officials A health authority or government official has said it, and no official has denied it.
- Several independent sources Two or more independent original sources back it. Outlets that only repeat the same original source count once.
- Only one original source Everything traces back to one original source, however many outlets repeat it.
- No source yet Someone told us, but there is no link to back it up yet.
- Contradicted Believable sources say the opposite.
- False We have established that it is not true.
Independent sources, not outlets
For every piece of evidence we record the original source the outlet relies on: an official (Rospotrebnadzor, the governor), another outlet (Lyudi Baikala, REN TV), an expert, or the outlet's own reporting. Chains are followed to the end: an outlet citing The Moscow Times, which cites REN TV, counts as REN TV. When five outlets repeat one Telegram post, that is one independent source, and a repeat is never more reliable than its origin. The one exception is official statements ("Rospotrebnadzor said…"): each outlet that reports the statement is a separate witness that it was made. Thesources page lists every original source and how many outlets repeat it.
The percentage
Each claim type starts from a prior: 25% for most types, and 15–20% for rumors, supply shortages and origin hypotheses. Every independent publisher that supports the claim moves the log-odds up by 2.5 × its reliability. Each additional publisher counts for 60% of the one before, because outlets copy each other. Contradicting publishers move it down in the same way. Several articles from the same outlet count once, and so does every unsourced tip combined. The result is clamped between 2% and 98%. An analyst can override either score and must leave a note when they do.
Bands: ≥85% very likely · ≥65% likely · ≥40% uncertain · ≥20% unlikely · below that very unlikely.
Source reliability priors
| Source type | Default reliability |
|---|---|
| official health authority | 90% |
| official government | 75% |
| scientific | 85% |
| wire service | 80% |
| established media | 70% |
| independent media | 60% |
| regional media | 60% |
| aggregator | 50% |
| organization | 60% |
| expert | 55% |
| tabloid | 30% |
| telegram channel | 35% |
| social media | 25% |
| ai agent submission | 20% |
| anonymous tip | 15% |
Known outlets have individual ratings, listed on the sources page.
Where the data comes from
- Automated collection. A scheduled job searches news (Russian and English), public Telegram channels and official sites. A language model breaks articles into claims, links them to existing claims, and records quotes.
- Submissions. People and AI agents can submit hints, evidence and symptom reports through our MCP endpoint. They are logged with provenance and reviewed before publication.
- Analysts review the queue, fix locations and set overrides when they know more than the formula does.
History
Every score change is recorded with the time and the reason (new evidence, a traced citation, an analyst decision). Each claim page shows its score over time, and the overview lists the biggest recent changes. History before 4 October 2026 was reconstructed by replaying the evidence in the order it arrived, using the current rules.
Trusted timestamps
When we first store a source, we save the exact text we read (a snapshot), compute its SHA-256 hash and have it signed by a free public timestamp authority (FreeTSA, falling back to DigiCert or Sectigo) under RFC 3161, the internet profile of ISO/IEC 18014. The signed token proves the snapshot existed, unchanged, at that time, even if the article is later edited or deleted. It proves what we captured, not that the publisher's claim is true. For search results without full text, the snapshot is the snippet we saw.
Each piece of evidence links to its snapshot and token. To verify one yourself (FreeTSA example):
curl -o s.txt https://plaguetracker.com/api/snapshots/ID.txt curl -o s.tsr https://plaguetracker.com/api/snapshots/ID.tsr curl -o freetsa-ca.pem https://freetsa.org/files/cacert.pem openssl ts -verify -data s.txt -in s.tsr -CAfile freetsa-ca.pem
"Verification: OK" means the file is byte-for-byte what was timestamped. Change one character and it fails. For tokens from DigiCert or Sectigo, use their root certificate instead.
Counts
Figures such as deaths or contacts under observation are cumulative totals as reported. The headline figure comes from the most likely report, and the range spans every report with a likelihood of at least 30%.
Symptom reports
Symptom reports are only published as aggregates per locality. Coordinates are coarsened to about 10 km, and free-text notes are never published. Each report gets aplague-compatibility score that combines fever with respiratory, lymph-node or skin signs and plausible exposure. The score only ranks reports for review and is not a diagnosis.
Limits
Official information from the region is limited and sometimes contradictory, and some outlets have strong editorial agendas. Treat every score as provisional. This site is not medical advice. If you have symptoms after possible exposure, call 112 or 103.