Developer builds deterministic Dota 2 draft solver after LLM underperforms at the game

A software developer built an offline deterministic solver for Model 322, a Dota 2 drafting game, after finding that a leading LLM performed worse at the game than he did himself. The LLM won only once in nine attempts, compared to a 14% title rate across 50 manual runs, though the developer noted the samples were too small and unpaired to serve as a proper benchmark. To enable fair policy comparisons, the solver used indexed random tapes so that competing strategies faced identical player offers, hero draws, and tournament conditions — an approach based on the common-random-numbers technique. The team also froze the game's client and data files with SHA-256 hashes and ported its rules into a verified Python engine to ensure a consistent definition of success. Finite opponent-field budgets were introduced to prevent any policy from gaming results simply by repeatedly redrawing easier matchups.