Prospective ledger
Live pre-event forecasts are immutable and content-addressed. Corrections append new outcomes and scores; nothing is overwritten.
Forecast trust
Evidence before claims.
Prospective forecast evidence is still being collected. Pick Pilot is not publishing an accuracy or calibration claim yet.
A forecast counts only when its receipt was recorded before the factual event, with a database-issued timestamp, model version, input hash, and exact source snapshots. Recomputed history is labelled reconstruction and never promoted to prospective evidence.
Factual scoring
Win, tie, and loss remain separate. We report Brier and log loss for probabilities, unit-labelled absolute error for numeric targets, and coverage only for predictive intervals.
No composite score
Eligible, matured, scored, void, missing, and corrected rows remain visible by source, league format, phase, surface, model, metric, and stage.
What is not evidence
Monte Carlo sampling-error bands are not predictive intervals. Producer self-confidence is not empirical calibration. AI narration quality is evaluated separately from forecast accuracy. A licensed-source credential does not activate a feed without accepted legal, security, and product reviews.