Product
Deterministic Replay: Watching a Decided Game Unfold
Novus Odds has no live feed by design. Here is how the replay tool turns a seed into a reproducible timeline you can scrub — with the result decided before the animation starts.
Key takeaways
- Replay has no live feed by design and never will — every fixture is generated from a seed.
- The result is settled by the performance model first; the timeline only visualises an already-decided history.
- Separating outcome from presentation is what stops the animation from smuggling information into the result.
- A seed is a complete, portable description of a fixture: share it and someone else sees the identical game.
- The win-probability trace shows what a market would have believed, not what the model already knew.
No live feed, on purpose
Replay does not connect to any live sporting event, and it never will. Every fixture is generated from a seed: the same seed always produces the same final score and the same timeline. That reproducibility is the point — you can share a seed and a colleague sees the identical game.
The design constraint is deliberate and it cuts in several directions at once. There is no real event to wager on, no feed to be wrong about, and no way for the tool to drift out of sync with reality — because it was never claiming to track reality in the first place. A synthetic fixture is honest about being synthetic.
It also makes the tool auditable. Anything you can see, anyone else can regenerate from the same nine characters of seed, on any machine, at any time. That is a much stronger guarantee than a screenshot.
The result is decided first
The performance model settles the final result up front. A separate timeline builder then distributes the scoring events across the game clock so you can watch it unfold, but the animation never determines the outcome — it only visualises one already-decided history.
As the timeline plays, the viewer recomputes an implied win-probability trace and the running edge, so you can see how a decided game would have looked to a market that didn't yet know the answer.
The ordering matters more than it appears. If the animation could influence the result, the outcome distribution would silently depend on presentation details — frame rates, event spacing, how long you left the tab open. By settling the score first and rendering second, the statistical properties of the fixture are fixed the moment the seed is chosen, and the timeline is a pure function of that score.
What determinism buys you
Determinism is not a performance optimisation; it is what makes the rest of the site's claims checkable. Every simulation in Novus Odds is a pure function of configuration and seed, which means results are reproducible, differences between runs are attributable, and bugs are demonstrable rather than anecdotal.
The engine gets there with an explicit seeded generator — an FNV-1a hash of the seed string feeding a mulberry32 stream — rather than the platform's random source. Nothing in the pipeline reads the clock. That is why a fixture generated today is byte-identical to the same fixture generated next year, and why the test suite can assert on exact values instead of tolerances.
It also enables a debugging technique that is otherwise impossible: hold the seed, change one parameter, and every difference you observe is caused by that parameter. No noise, no re-running until it reproduces.
Reading the win-probability trace
The trace answers a counterfactual: what would a market have believed at each point in this game, knowing only what had happened so far? It is recomputed forward from the state on the clock, so it never consults the final score it is walking toward.
Two features are worth watching. The trace steps sharply at scoring events and drifts smoothly between them, because time running out is itself information — a one-goal lead is worth more at 80 minutes than at 20. And the size of each step shrinks as the clock winds down for the leader and grows for the trailer, which is the mathematical version of a game getting out of reach.
Because you already know the result, you can watch the trace be wrong with confidence and see exactly how wrong it was allowed to be. That is a difficult thing to experience in a real market and a trivial one here, which is the entire argument for building a replay tool that has no live feed.
Frequently asked
Does Replay use real fixtures or live data?
No. Every fixture, participant and scoreline is synthetic and generated from a seed. There is no live feed, no sportsbook connection, and no wagering surface of any kind.
Can the animation change the outcome?
No. The performance model settles the final result before any timeline is built, and the timeline builder distributes events across the clock to match that already-decided score. Playback is presentation only.
Will the same seed always produce the same game?
Yes. Fixtures are pure functions of configuration and seed, generated from an explicit seeded RNG that never reads the clock or the platform random source. The same seed reproduces the same fixture on any machine, at any time.
What is the win-probability trace showing?
What a market would have believed at each moment given only the game state so far. It is computed forward and never consults the final score, so you can watch it be confidently wrong about a result you already know.
Try it yourself
Everything in this article is something you can run and tweak in the lab — with your own settings and a reproducible seed.
Open the lab →