This week's best bets.

The board above stays deliberately small because the model only plays lanes it has graded itself into. Its game view starts from FVOA-style team ratings — per-play offense and defense grades, opponent adjusted, garbage time discounted — which produce a line for every game; a play only makes the card when the model's number and the market's disagree by more than the noise band. On selective NFL moneylines that process backtests to +3.6% ROI net of vig, walk-forward (each season bet with only prior information), which is the number behind the homepage claim — a backtest, and labeled as one.

Why some markets are missing

Totals are gated off the board entirely: the model has never graded a durable totals edge, so it doesn't pretend to have one. Most weeks most games are no-plays — the NFL market prices the obvious things well, and what it prices correctly is exactly where a disciplined model declines to bet. QB props stay on the card because the projection they hang off — the same odds-implied projection surfaced on the rankings board — beats a walk-forward baseline across 7,000+ backtested player-games.

Graded in public

The backtest section above shows the record by season, not a highlight reel; whether books sharpen as the season runs is part of that research. During the season the weekly card is committed before kickoff and settled where it landed — wins and losses both stay on the page.