High-risk Stock trading needs to meet the above criteria for strict compliance and stable long-term returns. However, the existing methods generally combine market information and compliance rules in a single probabilistic retrieval-and-generation process, thereby weakening the constraint effect of compliance during long-chain reasoning and causing cascading error amplification. FinGuard is a multi-agent system that can reduce the number of high-risk stocks through the use of knowledge stratification. First, the continuously learning discriminator dynamically identifies and directs knowledge of different natures. Constraint-as-Code transforms rigid rules into executable logic, and can thus perform more explicit and consistent rule evaluation in the current system. A sliding safety audit mechanism continuously observes intermediate states in long-chain reasoning to reduce error accumulation and error propagation. In a controlled stock-trading evaluation over 1,764 trading days, FinGuard has shown stable results at the levels of day, week and month. FinGuard is better than the ReAct baseline and has increased compliance by 16.7% and adjusted returns by 14.7%. Stress tests show that both the drop and fluctuation are significantly smaller; therefore, FinGuard will meet the conditions for strict compliance and stable finances.
