How CrisisBrief collects

Which publishers’ reporting our forecasts rested on, and how those forecasts turned out. Each forecast is credited to the sources its own judgment cited, split evenly between them. A story copied by many outlets counts once, for the outlet that first ran it, and one an outlet carries from a news agency (named in its dateline or opening lines) counts for the agency. A publisher’s skill is how much better its forecasts did than always forecasting the average outcome. It is association, not accuracy: a source is credited for the forecasts that cited it, whatever its report said.

0 forecasts closed before 2026-07-29 and resolved, in 0 event families; 0 credited to a source. Skill of all of them: —.

Scores withheld. The independent check on how our forecasts are resolved is not passing (check 8: failed: volume: 1 labelled event families (1 forecasts); need 100; coverage: no forecast's window closed more than 30 days ago; audits: 0 of 1 needed on resolved_llm by a trusted auditor (no auditor is trusted on its market checks)), and these scores are learned from those resolutions.

A forecast counts 75 days after it closes: an event that happened is confirmed sooner than one that did not, so recent resolutions lean towards yes. The interval treats forecasts about the same event as one piece of evidence, and allows for the average outcome being an estimate too. Shrunk pulls a publisher with few effective event families towards the average, as if it had 10 more at the average. Where the brief ranks claims by source, it uses the shrunk score, and only for publishers scored here.