Track record.

Previous engine (Feb–Apr 2026, retired)

The record before the pipeline was rebuilt — reported per strategy version, like a fund’s track record. These forecasts stay on the books forever; they are never blended into the current engine’s headline numbers above, and never taken down.

A two-week transition build (Jul 1–14, 2026) sits between the engines: 2 resolved forecasts — far too few to chart, disclosed here so no era is unaccounted for.

Methodology

Every score is paired against the market: the Oracle’s Brier is only ever shown next to the Brier of the market price at the moment the forecast was made, on the same questions. Unresolved forecasts are scored against later market prices (closing-line value), with a terminal-regime exclusion — nothing is scored inside the final stretch of a market’s life, where prices collapse toward 0/1 on the clock rather than on skill — and a horizon cap of 25% of remaining life. All intervals are 90% bootstrap CIs (10,000 resamples); the headline p-value is a Monte-Carlo null that treats the market price as the true probability. Exclusions are counted, never silent: of 272 forecasts, 194 scored at 72h (1 terminal-regime, 29 horizon-capped, 26 without a matched price, 22 missing fields); 21 markets are additionally flagged and reported separately for possible self-impact.

Engine eras are attributed like a fund’s strategy versions: headline numbers come from the current engine only (since Jul 14, 2026); the retired previous engine is reported separately above with its own n and CIs. The two are never averaged into one flattering number.

Where n is small, the numbers say “collecting” rather than claiming significance — a good story at n=20 is still a story. Full definitions, guards and the reasoning behind each metric: EVAL_METHODOLOGY.md on GitHub ↗.

computed 2026-08-25T10:04:08+00:00 · tau = 72h