Sports Sim Battles

How sims work

Sport-by-sport engines step through real game states. Transition rates come from real world historical data, adjusted by player skills and (where applicable) coaching decisions. It’s deterministic for a given seed, so a rerun is the same episode. We also check that scores, pace, and margins look like real seasons, not video-game chaos.

Play-by-play, start to finish

Every battle is a complete simulated game told as a play-by-play script, the same format The Tape Room replays.

League-specific engines carry the load. Each sport runs a state machine matched to how that game progresses (for example, football as a drive and play-state model, basketball as an alternating-possession model), with the same pattern sized for more leagues. Every step is conditioned on the live situation, then samples the next play call and outcome from transition tables fitted on real historical play-by-play. Those tables follow a Markov-style game-state logic: given this state, sample the next event at frequencies grounded in league history, so skillful rosters still play inside real sporting shape.

Athlete and unit strength enter as rate snapshots predicted before the battle with gradient-boosted models, then clamped into the engine so skill moves the needle while calibration holds. Inference runs at lock time; what you locked is what the sim sees.

Seed policy

  • Sim seed drives game outcomes (including venue and crew draws where the sport uses them). Same sim seed produces the same result tape.
  • Presentation seed shapes Tape Room flair only (for example shot placement chrome). It stays on a separate stream from outcome math.

Golden tapes and regression tests lock that reproducibility. Offline accuracy packs and Monte Carlo-style seed sweeps pressure-test win rates and margins against the corpus; each live battle is one faithful seeded game.

Calibration & vintages

Engines are checked for score, pace, and margin shape against real-season targets. Every completed battle tape records the calibration version and the ratings / model vintage that were locked with the entry. When those stamps change, new battles pick up the new stamp; already-played tapes keep the vintage they were simulated with.

Fair outcomes

Strategy is drafting the right athletes. Cap salaries and Battle Points stay outside who wins. Drafting does the work. Versus and Showcase modes follow the published mode rules; the engine’s job is fair, reproducible, rules-aware sport.

Scope of this page

This page covers public methodology: seeds, calibration, and how the play-by-play engines behave. Model weights, training data, AI feedback, and other proprietary information is kept internal.

Also useful

Play