They said it was inevitable. We wrote it down.
240 dated predictions from 209 of the people building AI — quoted verbatim where a source exists, summarised where it does not, and graded only after the horizon passes. This is the ledger, not the prophecy.
Twenty verdicts are in. Thirty-six clocks are still running.
Of 56 sourced predictions, 36 have not reached their horizon and are therefore not scored — not "probably right", not "likely wrong". The remaining 149 records are editorial summaries: our reading of a widely argued position with no primary citation yet, so they are marked unrated and excluded from every total below.
The rule that does the work
Everything else follows from that: why two thirds of the sourced record is unevaluated, why an unfalsifiable claim is recorded but never counted, and why an editorial summary is labelled rather than graded.
Every number here is in one file.
The dataset ships with the page and is the source of truth. Original text and dataset are CC BY 4.0; linked sources remain the property of their publishers.