How we got there: how much of a week is real

We asked: how much of a single NFL week is signal about the player, and how much is dice?

Method: for every QB season with 10+ games of 15+ dropbacks (n=286), deal the games into odd and even piles, compute each stat in both piles, and correlate across players. Spearman-Brown converts the half-season correlation into single-game reliability. All inputs are public in the player_week view.

The published table (scripted; construction in the article backtest block): single-game signal fractions - rushing yards .43, aDOT .12, EPA/dropback .11, sack rate .11, yards/attempt .09, completion pct .09, CPOE .08, TD rate .08, and INT rate .013 (one game of interceptions is one percent signal; it takes 79 games to reach half).

Altitude is EPA per dropback with the luck deleted: interception plays re-priced as that season's average incompletion, sack-fumbles re-priced as ordinary sacks. The de-noising lifts single-game reliability from .109 to .128. An earlier composite that added rushing identity scored .123 against the plain stat and was retired; that autopsy has its own article. Bands, not decimals: Grounded, Turbulence, Level, Cruising, Soaring. Every published score is a row in the public qb_altitude_week view, raw and de-noised EPA side by side.

The v3 upgrade: an informed prior. The score now shrinks the de-noised EPA toward a decision-priced expectation of the same dropbacks, which carried reliability from .128 to .147 while validity rose from .505 to .518. The cell below runs the live association between that expectation and de-noised production; both columns ship in the public qb_altitude_week view.