Every NFL game, played 10,000 times before kickoff. The model's number next to the market's. Player projections from the same runs. Every figure timestamped, graded weekly, record public.
Three views of the same 10,000 runs. Nothing here is a pick; the record page says how the model has done, and against what.
Each game is played snap by snap: play calls, yards, turnovers, fourth-down decisions, kicks, clock and overtime, sampled from five seasons of NFL play-by-play. Team strength is opponent-adjusted EPA per play for four units, with the quarterback rated on his own.
The model's line sits next to the market's for every game, with cover and over probabilities counted straight from the runs. Where they disagree, you see by how much, and how that pattern has graded before.
Every number is timestamped before kickoff and graded against the result every week. Wins, losses and the market's score sit on the same page, in the open, back to 2022.
Weekly simulation outputs for every NFL game: lines, totals, player projections, fair prices. No picks, no narratives. Free while the record is being built.
Model line is the median simulated margin. Cover and over probabilities count runs, so key numbers are priced by the distribution. Lines are the nflverse schedule snapshot and can lag your book.
Projections from the same simulated plays that produced the scores. Targets and carries follow each player's share of recent usage; players off the active roster are removed. Fantasy points use the scoring you pick.
A prediction-market contract pays $1 if the team wins, so its fair price in cents is the model's win probability. Compare with what Kalshi or Polymarket is asking, after fees and against the book depth.
Exchange columns fill in when the Kalshi and Polymarket adapters are connected (both have public read APIs). Sportsbook implied probability is the no-vig moneyline from the schedule snapshot. Kalshi charges roughly 7% of expected profit per contract; a fair-price gap smaller than the fee is not an edge.
Walk-forward over 2022–2026: every projection uses only games played before that date, graded against the closing line in the nflverse schedule file. Tuning used 2022–24; 2025–26 is untouched holdout.
Break-even at -110 is 52.4%. The holdout rows are the ones that count. Read plainly: the simulation alone does not beat the closing line on sides or totals. The blend of model and market matches the market's accuracy. The model's value is the distribution it produces for every market the book doesn't publish a number for, and the fact that all of it is graded here.
158,000 scrimmage snaps from 2022–2026 become outcome records (yards, result, clock) bucketed by play type, down, distance and field zone, sorted by EPA. The simulator draws a quantile and shifts it by the matchup: offense rating plus defense-allowed rating, calibrated so a 0.10 EPA-per-play gap moves outcomes by 0.10. Ratings are opponent-adjusted, decayed weekly, shrunk toward league average, and strictly as-of. Play calls, fourth downs and field goals are fitted models; clock and special teams are empirical. League scoring is calibrated to 44.9 points per game. Home field is worth about 1.3 points. Prior-week market power ratings were tested as an input and did not help out of sample, so they are off.