Trust scores are evidence-weighted,not just hit rates.
Accura copies prediction-market resolution discipline (Polymarket / Kalshi): named venue, inclusive window, early Yes / late No, Too Early, and Unknown-as-void. It does not copy payouts or 50-50 refunds.
Write the contract first
A claim is only scoreable when it names an asset, a metric (high / low / close / event), a venue, and a window. The title is marketing; the contract is law.
Code grades, not the model
Path high can resolve Yes the day it prints. A miss waits until the inclusive window ends. Unmet conditionals and incomplete scraps are void — not half-hits.
Four layers
Coverage is not skill. Specific accuracy (named things) is the headline. Directional sides are separate and shown vs BTC. Calibration buckets “definitely” vs “could”.
Propose then challenge
Code proposes a grade with candle evidence. An undisputed proposal is accepted only after a 48-hour challenge window; a changed proposal restarts the clock. Too-early proposals are rejected. Disputed rows stay off the public score.
The person, not the channel
Every prediction is attributed to the exact speaker with a confidence. Guest calls count for the guest. Attribution below high confidence is reviewed by a human, never counted.
Brier and calibration beside the hit rate
Hedge words imply a probability (definitely ≈ 90%, could ≈ 55%). We publish the Brier score, Brier skill versus chance, and the calibration gap. A speaker 25 points overconfident on three or more calls cannot be Verified.
Headline formula
Only resolved specific contracts (threshold, close, floor, target, event) enter the number. Confidence comes from language: definitely = High (1), I think = Medium (0.7), could = Low (0.4). score = round(accuracy × min(1, weightedResolved / 8) × 100). No number before 3 resolved claims; verified needs 8, full evidence weight, and speaker attribution. Void and pending stay out of the denominator. We also publish a Wilson interval and shrink toward 50% until the sample is large.