Truco-Bench: How Rules Shape Signaling Between Language-Model Agents

An agent’s use of a communication channel depends on the rules it is told to follow. We introduce Truco-Bench, a deterministic partnership card game for studying this dependence while keeping channel mechanics fixed. Two language-model agents hold private cards and can send messages visible to their partner and an external monitor, but hidden from opponents. Across Qwen3-8B, 14B, and 32B, with 100 hands per condition, no channel use occurs when signaling is described as forbidden and no explicit request to signal is given. Describing it as allowed produces use in 86–99% of hands; explicitly requesting signaling restores use under the forbidden framing to 96–100%. Messages convey limited private-hand information, without an established improvement in game outcomes. Prompted message monitors weakly distinguish signaling-policy from nominally honest- policy messages. A separate exploratory activation probe predicts whether an agent emits a message; it does not establish detection of covert intent. These results characterize instruction-sensitive signaling and expose the measurement controls needed to evaluate agent communication under oversight.