Hit-rate — for each recorded snapshot where a voter called a direction (|lean|>0.05), did price move that way by the horizon? Hits ÷ calls. 50% = coin flip; the edge is hit-rate − 50%.
Suggested weight — scales with the voter's aggregate edge across horizons (50%→0.6, 60%→1.0, 65%→1.2; below 50%→down). Replaces the heuristic Oracle weights when applied.
Reliability curve — the real test of probability: bucket every P(up) prediction, plot predicted vs realized up-rate. On the diagonal = honest probabilities (the Phase-3 goal: "70%" means 70%).
Accumulation — built from the recorder (every 5 min via cron). Near horizons calibrate in a day or two; 1-week hit-rates need ~2 weeks of history.