Open data
Audit the record
The forecast record is published as plain machine-readable data. Download it and compute the Brier scores, calibration and anything else yourself — the summary pages have no information the raw data doesn't.
Downloads
| File | Contents | Format |
|---|---|---|
| forecasts.json | Full forecast ledger: typed distributions, update history, outcomes | JSON |
| cohorts.json | Ex-ante cohort registry (Core / Featured) and coverage membership | JSON |
| commitments.json | Public sealed-forecast commitments (hashes and reveals) | JSON |
| forecasts.csv | Flat view for spreadsheets: one row per forecast state (every update), with quantiles or probability, outcome, scores and the primary benchmark for the standing state of resolved forecasts, and the earliest independent capture of each state. Derived from the JSON ledger and regenerated with it | CSV |
| archives.json | Independent Internet Archive captures of the ledger, each SHA-256-matched to the published file, with the forecast states each capture contains. Re-check any entry by downloading its raw_url and hashing it | JSON |
The JSON is the canonical source; the CSV is a derived view and is checked against it before every publication. A documented schema will be added as the record grows.
Scoring, reproducibly
Forecasts are scored with proper scoring rules — a logarithmic score across all types, with Brier and CRPS as secondary diagnostics, and each forecast measured against an uninformative baseline. Every aggregate on the track record is derived from these files in your browser by the same open scoring code — so if your computation disagrees with the site's, the site is wrong. That is the intended failure mode.