Weekly Brier score + Murphy 1973 decomposition (Reliability − Resolution + Uncertainty) on the binary claim “classifier-bullish ⇒ beats SPY at 7d.” Reliability diagrams use the CORP method (Dimitriadis-Gneiting-Jordan, PNAS 2021) — isotonic regression replaces ad-hoc equal-width binning.
Ship gate: Brier ≤ 0.24 with |base_rate − 0.5| < 0.1 (T-20-C-02-01). Minimum n = 100 per classifier_version (T-20-C-02-02). Recomputed Mondays 08:00 UTC via /api/cron/eval-brier.
See also: /insights/sentiment-health · /insights/sentiment-sources
No Brier evaluation written yet.
The first run is scheduled Monday 08:00 UTC via /api/cron/eval-brier. Manual runs: tsx scripts/eval-brier.ts.