Three layers, three kinds of evidence
- Historical start
- The recorded score, time, possession, field position, down and timeouts at a chosen moment. Each published challenge links to its source.
- Simulation rules
- The legal actions, clock transitions, scoring and end conditions supported by this game version.
- Outcome model
- The possible gains, incompletions, sacks, turnovers and durations assigned to an action. A plausible branch is not the real next play.
Corrections to a source fact and changes to an outcome model are different changes. Version labels make that difference visible.
Scope of the drive
The simulation follows one offensive possession and any supported try after a touchdown. It does not model the opposing offense's response, a full game's final win probability, injuries, detailed formations or every officiating exception. Only the actions and rules exposed by the current version are available.
The clock advances through modeled setup and live play. It does not run while you read or decide. A long gain tackled inbounds can leave preparation time before the next snap.
The empirical outcome model
The broad run and pass samples come from penalty-free plays in the final two minutes of the second and fourth quarters in the 2023 nflverse play-by-play release. The 2024 season is held out for descriptive checks. This is a pooled model, rather than a separate model of each player, team or defense.
Gains, event type and boundary finish are sampled together. Sack risk is pooled across pass calls because the intended route depth is not observed on sacks. Live-play durations and the 10-second ordinary setup are explicit design assumptions. A running-clock spike uses 6 seconds to set up and 1 second live.
See nflverse attribution and source links. Data validation does not turn these game assumptions into real-world forecasting evidence.
Seeds and replay
A seed makes an outcome sequence reproducible when the challenge revision, rules, model and action sequence are also the same. The seed alone is not a promise that different actions will produce equivalent luck.
Replays display recorded events. Animations and sound describe those events without changing the outcome. Old records may remain viewable even when their versions are no longer comparable with new attempts.
Scoring and decision feedback
The drive result tells you whether the challenge's stated objective was met. Resource changes explain where seconds, downs and timeouts went.
Open “Review the last decision” for a comparison of the available calls. Each call is evaluated through 128 bounded simulated continuations under a fixed contextual policy. The displayed percentage is the share of those simulated possessions that meet the objective. It is not a measured NFL win probability or a ranking against real coaches.
The 95% intervals describe sampling uncertainty inside that experiment. They do not include errors in the timing assumptions, pooled outcome model or continuation policy. When the intervals do not separate the selected call from the strongest alternatives, the game reports an uncertain comparison and withholds a numerical quality score. When they separate, the score compares the selected modeled completion estimate with the highest one. A fortunate result alone does not grade the decision.
What a fair comparison requires
Compare attempts only when their challenge, mode, start, rule and model versions match. First Daily attempts and later practice replays serve different purposes. This release does not claim a competitive ranking or population percentile.
Sources and corrections
Use data sources to inspect the reference material. Changes that affect saved attempts belong in release notes. To report an error, include the challenge and version shown on the result screen through Support.