The Proving Ground — what can be proven before July 2027, and what cannot
Every sealed forecast on the ledger is still pending, so the site claims no track record. That does not leave a reader in the dark. The first items to be graded are two scored public claims dated the end of 2026 — verdicts in January 2027 — followed by the first sealed forecast on 28 February 2027 and the first public Brier scores in July 2027, under a scorecard written while every seal was still pending. Two easy, low-probability calls resolving first proves that the grading procedure runs as written; it does not prove skill, and the page says so in advance.
Three tracks of evidence
Track C, the prospective record, has been running since September 2026: numbers written down before the outcome, plus a nightly log of the live lean on every seal and the raw reading before the ±7 rail. Track B, planned: the Brier score that purely mechanical base-rate rules earn on public, timestamped question sets — the bar the July 2027 score has to beat. Track A, planned and to be pre-registered first: whether the engine's thirty-six strands measure anything, by running archived headlines through a frozen lexicon and comparing each strand with an index somebody else built — financial stress, geopolitical risk, policy uncertainty, global temperature — with quiet years as a placebo.
What will not be published
A backtest against the site's own historical curves, which are a stylized model curated with hindsight. Forecasts "as of 1999" written by a language model that has read how it ended. Retro numbers blended into the Brier record. A retro-test tuned until it looks good — backtest overfitting. Nothing on the page is a result: it is the plan, dated and published before the work, and results will be added beneath it rather than in place of it. See how a seal is resolved.