SGSzero: Training a Compact Sanguosha Agent via Privileged Rule-Based Supervision

We present SGSzero, to our knowledge the first decision-making agent for Sanguosha, a multi-player Chinese card game with several hundred hero skills, chained effect resolution, and hidden hand information. Deploying a learned agent for such a game on consumer hardware imposes two constraints rarely studied together: the agent must operate under information available to a human player, and it must run within a latency budget that precludes large language models. Our central observation is that a handcrafted rule engine — which has access to internal card valuations invisible to a human player — can serve as a privileged teacher in the Learning Using Privileged Information (LUPI) framework. Training a compact Transformer (1.76M parameters) with a decision head plus an auxiliary card-value head, then discarding the value head at deployment, yields 80.9% decision accuracy on the "play or pass" sub-task while consuming only the state a human player can observe. Naively exposing the privileged valuations as input instead causes a deployment failure (83.9% in-distribution, 35.3% out-of-distribution), which we diagnose and attribute to the model over-relying on numeric features. We further report three empirical findings: (i) model capacity is not the bottleneck — seven hyper-parameter configurations all plateau at 72-73%, whereas a 7.6% increase in data raises accuracy by 5.3 points; (ii) 89% of remaining errors fall below 0.9 confidence, boundary calls rather than systematic misconceptions; and (iii) a local collapse masked by the aggregate metric is repaired by the privileged value head, improving that subgroup by 17.2 points. The resulting model trains in 70 seconds on a single consumer GPU and performs inference on commodity CPUs. Code: https://github.com/wenzer0/SGSZero