AWS solves NHL playoff clinching with CP-SAT and a custom tree search — no LLM in the loop

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

The AWS Generative AI Innovation Center built a two-part solver that proves with mathematical certainty when an NHL team has clinched: a 0-day constraint-programming model on Google OR-Tools' CP-SAT that tests whether any scenario exists where the team misses, and an n-day lookahead using a custom tree search where each layer is a game day and each node an outcome, calling the 0-day solver at every node. Median runtime is on the order of minutes for 1-day scenarios with pruning efficiency near 100%, validated against four NHL seasons (2021-2025) where output matched the league's published scenarios exactly. Worth noting for builders: the problem — seven cascading tiebreakers requiring exact verification — is precisely where a probabilistic model is the wrong tool, and a team inside a generative AI org said so.