Large language models can produce fluent yet factually incorrect responses, limiting their reliability in knowledge-intensive applications. Recent decoding methods seek to mitigate such hallucinations by contrasting or amplifying signals from intermediate layers. However, not every layerwise discrepancy provides a reliable direction for factual decoding. We propose a unified game-theoretic decoding framework for large language models: intermediate-layer signals act as competing evidence, while the final layer retains decision authority under explicit constraints. Rather than applying a fixed layerwise contrast, our framework treats changes in intermediate predictions as evidence of the language model’s evolving internal judgment, and uses this evidence to guide decoding under final-layer constraints. Experiments with three backbone language models on multiple truthfulness and hallucination benchmarks demonstrate the effectiveness of our framework in improving factual reliability.