Janus-Q is a two-stage end-to-end event-driven trading framework that elevates financial news events from auxiliary signals to primary decision units. In Stage I, the authors construct a large-scale dataset of 62,400 Chinese financial news articles, each annotated with 10 fine-grained event types, associated stocks, sentiment labels, and cumulative abnormal returns (CAR) computed via market model and risk model neutralization. This dataset addresses the lack of joint modeling of event semantics and market impact magnitude. In Stage II, Janus-Q performs decision-oriented fine-tuning: first, supervised fine-tuning (SFT) establishes a reasoning-aware mapping from event descriptions to expected CAR; then, reinforcement learning via Group Relative Policy Optimization (GRPO) is guided by a Hierarchical Gated Reward Model (HGRM). The HGRM decomposes the trading reward into interpretable components—direction correctness (hard gate), event-type consistency (soft gate), cost-aware profit/loss, magnitude shaping, and process rewards—to align language model reasoning with financially valid trading behavior. Extensive experiments on a test period from November 2024 to February 2025 show that Janus-Q consistently outperforms market indices, time-series-oriented LLMs, financial domain-specific LLMs, and general-purpose LLMs. It achieves a Sharpe Ratio of 1.3088 (102.0% improvement over the runner-up), direction accuracy of 58.69%, and event-type accuracy of 80.09%. Ablation studies confirm the critical role of each component, particularly SFT and the diversified reward objectives, in achieving robust and profitable trading decisions.