The Wnuan framework introduces a three-stage post-training pipeline designed to integrate proprietary enterprise knowledge into large language models while maintaining general capabilities. The methodology involves constructing task-oriented supervision directly from documents, executing supervised fine-tuning (SFT) paired with general-data replay, and employing reinforcement learning (RL) to address residual errors. In evaluations using the 707-question WnuanBench, a 32B parameter model demonstrated significant performance improvements, increasing its acceptable-answer rate (AAR) from a baseline of 52.76% to 80.06% following SFT, and reaching 91.51% after the RL stage. To optimize the RL phase, the researchers implemented a residual-error sampling protocol, which outperformed full-pool and size-matched random sampling by 3.11 and 2.97 points, respectively. While the adaptation successfully internalized domain-specific data, it incurred a performance cost on general benchmarks, which saw an average decrease of 5.17 points, primarily affecting instruction-following abilities. Validation results showed high reliability, with an automatic evaluation ensemble achieving 90.5% agreement with authoritative domain experts on stratified response samples. This research provides a structured approach to enterprise-specific model adaptation, quantifying the trade-offs between specialized knowledge acquisition and the preservation of general reasoning and instruction-following skills.