Get Reliability

The receipt: does a published hold probability mean what it says?

Returns an overall block plus one per horizon, each carrying the Brier score, the base-rate baseline it must beat, the sample-weighted calibration error, and the reliability table behind those numbers.

Per-horizon matters as much as overall. The term structure is the part of this model with an explicitly empirical constant in it (CONE_TERM_DECAY), so a systematic miss concentrated in the +2h bucket is the signal that the constant is wrong — and it is visible here before anyone has to argue about it.