We grade how far a report can be traced — never whether it's true.

Sources

Every public case comes from one recorded source report. These are the sources behind the cases on the site today:

Counted from the database when this page is built (refreshed at most hourly). Only cases a reader can open are counted.

Some archives let us say that a report exists but not show its text. Those cases are pointers: the facts we may state and a link to the source.

Grading: Reported → Searched → Verified

Each case sits at one step. The step says how far we could trace the circumstances. It says nothing about whether the phenomenon is real.

• ReportedSomeone reported it, and that is all we claim. No search for corroboration has been concluded. 4 363 cases
• SearchedWe searched for independent sources and found none that qualify. The case stays on the record as reported. 3 639 cases
• VerifiedAt least two independent carriers of the same event: the original report, plus at least one source that is independent of the person telling the story. That confirms the circumstances, never the phenomenon. 7 cases

Provenance: P1–P4

Separately from the step above, every case carries a provenance grade: how close the record is to the original witness.

• P1First-hand: the person it happened to, in their own words.
• P2Someone retelling a first-hand account.
• P3Reported by press, researchers or catalogues.
• P4The trail is unclear.

Both scales grade traceability. A P1 account that is Verified can still be mistaken; a P4 report that is only Reported can still be accurate.

How "independent" is decided

A source counts as independent only if it was produced by someone other than the person telling the story — third-party journalism, a reference work, someone else's study, a court or registry. The witness's own site, channel, book, social post, or a repository relaying their own material does not count, whatever their title.

• Independence is checked structurally in code (who published the source, whether it is the case's own origin, whether it names the people and the event). An AI model may only downgrade a source in this step; it can never make one count.
• A case also needs a circumstance that can be checked from outside the report itself. Without one it stays Searched.
• Press coverage alone counts only when at least three distinct outlets, none of them the origin, carry the same name, place and event.
• The judging step is run five times; Verified requires at least four of five. A split falls to Searched — doubt falls downward, never upward.
• A verdict that a source contradicts the report is never written automatically; it is held for a person to read.
• Many people seeing the same thing at once is marked as one collective account, not as corroboration.

"N independent outlets" on a case is the number of distinct independent sources found beyond the origin. It is not a vote count.

Where AI models are used

The models are run locally on our own hardware. The main one is Qwen 2.5, an open-weight language model with 7 billion parameters. Steps that use it:

• Extraction. Reading a forum post, podcast transcript or article and pulling out people, places, dates, institutions, the phenomenon type and a one-line description of what was reported. Registers and catalogues with structured records are parsed by ordinary code instead.Model: Qwen 2.5 7B · Review: no person reads each case before it appears
• Type check. A second opinion on the phenomenon type. A case is moved to a new type only if the model proposes it and the text contains a keyword for the new type; otherwise it stays where it is.Model: Qwen 2.5 7B plus keyword rules · Review: automatic relabels are spot-checked by hand on a sample before they run at scale
• Translation. Non-English source text is translated into English for display; the original stays linked.Model: Qwen 2.5 7B · Review: none per case; translation accuracy has not been measured
• Summaries. For press and other published reports, a short summary written as "according to the accounts". First-person forum and podcast accounts get no composed summary: they are shown as quotes with a link, under the one-line description.Model: Qwen 2.5 7B · Review: none per case
• Corroboration search. Writing search queries, judging whether a found source is about the same event, and flagging sceptical or debunking sources.Model: Qwen 2.5 7B · Review: the model can only downgrade; the decision that a source counts is made by the structural check above

Corrections, duplicate decisions on the same document, and "contradicted" verdicts are made by a person reading the sources.

Duplicates

• The same document twice (for example the same paper via Europe PMC and PubMed) becomes one case. The other copy is held with a note saying which case carries it. These pairs are read by a person.
• The same event from different publishers stays as separate cases, linked as "Also reported by". Two unrelated publishers on one event are corroboration, not noise.
• News items that share an event key (built from a named person, institution or place) are merged automatically into the existing case.

Measured error rates

• Phenomenon type. 30 random public cases were read blind by a person on 12 September 2026; 29 could be judged. The type shown matched the reader's in 19 of 29 (65.5%), or 21 of 29 (72.4%) if "Unclassified" is not counted as a miss. A re-run the next day gave 20 of 29.
• Invented person names. On 18 September 2026, among 457 public cases with person names where we know exactly which text the model read, 0 carried a name that does not appear in that text.
• Automatic news merges. Hand-read on 26 September 2026: in 38 of 68 judgeable cases (56%), all merged articles were about the same event. The rest mixed different events or could not be decided.
• Verified. How often a Verified case is wrong has not yet been measured. Neither has translation accuracy or the accuracy of summaries.

Corrections

When a case is wrong, we correct the specific field against the case's own stored sources. Anything the sources do not settle is left unchanged rather than guessed, and the previous value is kept so the correction can be reviewed. When a verdict is withdrawn or restored, the case's detail view records it in that case's own history.

Getting the data

• A read-only JSON API, no key: /api/search, /api/case/<id>, /api/verified, /api/flow, /api/stats.
• /llms.txt lists the key pages and endpoints and /llms-full.txt every public case as one line of facts; /sitemap.xml lists every public case; RSS at /feed/all.
• Reuse terms: /license.

Visit counting

Each page view sends one anonymous beacon with the page, the referring site's host name and any utm_source in the address (plus device class, country from our CDN, and time on a detail view); it sets no cookies and stores no IP address. Details: /privacy.