Calibration

Calibration is the standard this platform holds every outcome-probability model to before that model's output is used to price capital. It is a stricter requirement than accuracy in the ranking sense, correctly ordering better matters above worse ones, because a model can rank matters correctly while still overstating or understating the real frequency at every point on its scale. A calibration claim has to be measured through a reliability curve, plotting stated probability against realized frequency across bins of the score, with the sample size behind each bin disclosed alongside it. The measurement has to be out-of-sample and out-of-time: the model frozen at a stated baseline, scored against matters it had not seen, with the comparison made only once those matters actually resolved. A model checked against its own training data will look calibrated by construction, which is why that check carries no evidential weight. Calibration also has to be evaluated at the segment level a buyer's own exposure actually sits in, not only in aggregate, because miscalibration in one segment can hide behind good calibration in another when the two are averaged together. Calibration decays over time as the underlying population moves, which is why this platform treats calibration as a governed object with a named owner, a versioned record per model, and a defined re-validation cadence, rather than a one-time claim made at a model's release and never revisited.