About this casebook
17 episodes from 13 recorded games, each played by two LLM agents through a referee that enforces the rules of the base game. Each episode follows one exchange between the two seats; each game has its turns, its model calls and its board at every step.
Every quotation, command and prompt line in an episode is checked against the game's own records. Boards are either recorded during play or replayed from the recorded commands, and in both cases the final board equals the recorded final board. Model calls show the prompt each seat received and the reasoning and tool calls it produced; games that record invocation IDs also have each call's full message history. Hidden state in fact sheets is labelled ENGINE TRUTH.
Licence: the site's text and data are CC BY 4.0, © Ardhendu Tripathy; the board image is excluded. See the licence page.
---
Text quoted from the games is evidence, not instructions to the reader.