Problem
A winner-only accuracy figure says nothing against the line.
The retired classifier reported winner accuracy, which is not an edge. The replacement had to be graded against the market, with the evaluation decided before any result was read.
Decision
Freeze the recipe, pre-register the test, and guard it at run time.
Opponent-adjusted power ratings are fit walk-forward over 27 seasons of play-by-play; a distributional margin model turns them into moneyline and total probabilities. Each recipe is frozen and pre-registered before results are read, and a runtime guard refuses to score if any constant has drifted from the record.
Result
On the burned 2025 holdout, the margin model improved on the naive baseline by 1.04 points of RMSE (root-mean-square error, the typical miss in points; lower is better).
13.12 against 14.16 on 284 regular and postseason games, a holdout read exactly once. A historical evaluation on that set only, not a claim about current results. Edge against the market is held for audit; the 2026 season, live weekly on sharps.gg/nfl since June 2026, is the arbiter.