HardRead

I play competitive Pokémon, and after a loss I never know whether it was the crit or me. HardRead helps players investigate that question. Paste a finished Gen 9 OU replay, review the turns that changed the game, and ask questions about the result or your own read of the opponent. The analysis comes from Foul Play. It samples possible opposing sets and searches for good moves each turn. HardRead adds estimates of luck, a breakdown of the result into starting position, luck, and choices, and a way to test the player’s stated read. Meta’s Muse Glimmer-30B explains the engine’s findings. Rather than receiving the replay in one prompt, it chooses read-only tools to investigate the game. It writes sentences with placeholders linked to engine results, and code inserts the values. A checker rejects numbers without those links and citations to the wrong type of value. Players can inspect every lookup and rejected draft. That makes numerical claims traceable, but it does not guarantee a correct explanation. The checker can miss a number taken from the wrong record, and it does not check claims expressed only in words. In an evaluation of three questions using the 28 October versions, before a fix for hindsight about hidden sets, 41 of 60 answers passed the checker immediately and 59 passed after one repair. A blind AI review flagged 3 of 60 sampled sentences; no human has validated the answers. Separately, a 10,000-game comparison of our engine change against Foul Play produced a 50.05% win rate (95% CI 49.16–50.94), so we do not claim an improvement over it. Luck and win chances are estimates and can vary between runs. “Choices” includes both players’ decisions and model error; it is not a skill score. We have not run a study with human players. Try the six saved reviews, which open instantly and include the saved Glimmer answers. Start with showcase 1 for a game where estimated luck swung the result, or showcase 5 for one where choices made the difference. When live analysis is online, you can also submit a finished Gen 9 OU replay. I designed and directed HardRead; AI coding agents, including Claude Code and Codex, wrote much of the code.