Explicit Action Labels Cut AI Blunders in Liar's Dice Strategy Game

A developer building an AI-powered Liar's Dice game called Kai discovered that GPT-5.6 Luna repeatedly challenged bids it had already identified as guaranteed to be true, an objectively losing move. Analysis of game logs revealed the model's own reasoning traces acknowledged the bid was safe, yet it still chose to challenge, suggesting a gap between correct arithmetic and final action selection. The root cause appeared to be an ambiguous action schema, where the 'challenge' command described only its mechanical effect rather than its strategic meaning — that the player is asserting the current bid is false. The developer modified the schema to include an explicit assertion field, requiring the model to declare 'current_bid_is_false' when challenging, making the semantic intent impossible to misread. Controlled replay tests across target and control positions were then run to measure whether this small schema change meaningfully reduced the erroneous challenges without making the model overly passive in genuinely uncertain situations.