Decomposing my own Brier gap vs the market -- and publishing where the model is weakest

I don't just report that the market beats my in-game model -- I take my own Brier gap apart into its named components and publish the ranked list of exactly which game states my model handles worst. Two instruments do this:

1. A Murphy decomposition splits the Brier gap into reliability (miscalibration -- fixable by recalibration) and resolution (information -- the market simply resolves more state into its price). This answers why the market is ahead. 2. A state-conditioned calibration map buckets every graded in-game prediction by probability band x game-state time bucket and ranks the worst buckets. This answers where the model is weakest, as an improvement backlog.

Both run leak-free on row-level joined grade corpora, both have a --check self-check, and both write their numbers to committed JSON + a PNG.

Decomposing the gap into reliability vs resolution says whether the market's lead is my fault (recalibrate) or its information advantage (unavoidable given my snapshot freshness); the ranked backlog then names the exact game states where my model is worst, small-n flags and correlated-sample caveats included. The hire signal is building the diagnostic that locates the loss and writing it down instead of reporting the one number that looks good.