Play the game before it's played

The engine resolves tonight's posted lineups and probable starters (or each team's most-used nine for a hypothetical), then pulls every batter's and pitcher's plate-appearance history from the corpus.

Each PA is an eight-way outcome draw — walk, HBP, strikeout, homer, triple, double, single, out in play — from the batter's rates composed against the pitcher's allowed rates, shrunk toward league averages so a hot week can't hijack the math.

Starters get a realistic leash and hand off to a bullpen composite; runners advance, double plays turn, extra innings happen. Every simulated game is a complete box score, not a coin flip.

Win probability, expected total and margin, the full score matrix, per-player stat lines and how often each happens. One number lies; ten thousand games tell you the shape of the thing.

√N convergence

The uncertainty on a win estimate shrinks with the square root of the number of games: 100 sims puts a ±10-point band around it, 10,000 sims about ±1. The convergence chart in every run shows the band closing in real time — you can literally watch the answer settle.

Distributions beat point estimates

“Dodgers by 1.4” hides everything interesting. The simulator hands you the full score matrix — how often it's a blowout, how often it goes extras, how often the underdog steals one — which is exactly the shape a market price is made of.

Two engines, honestly labeled

When lineups and starters resolve, the player-level engine (pa_v1) plays every plate appearance. When the matchup is too thin for that, a team-level engine (nb_team_v1) samples from each side's scoring distributions instead — every result tells you which engine ran.