OOM-RL: Out-of-Money Reinforcement Learning - Market-Driven Alignment for LLM-Based Multi-Agent Systems

Battle Count: 63

Highly relevant to quantitative trading. The paper directly addresses live equity trading (CSI 300), market microstructure friction, execution costs, portfolio construction, factor attribution (Barra-style), and risk management. The OOM-RL framework uses real trading PnL as the alignment signal. However, the primary contribution is in AI alignment methodology rather than novel trading alpha generation. The trading results (Sharpe 2.06, 34.48% annualized return in mature phase) are impressive but based on only 94 days with marginally significant alpha. The execution friction modeling (0.08% per transaction, slippage, liquidity constraints) is directly applicable to quantitative trading practice.

Very high implementation complexity. Requires: (1) Live market access with brokerage integration, (2) Docker container orchestration with OS-level RO-Lock enforcement, (3) AST-based code mutation pipeline, (4) ≥95% code coverage across 8,300+ lines, (5) Cryptographic state hashing for Byzantine fault tolerance, (6) Frontier LLM API integration (GPT, Claude, Gemini), (7) Real-time market data and order execution infrastructure, (8) Human expert domain knowledge for Epistemic Autopsy prompting, (9) 20-month longitudinal deployment for meaningful results, (10) Capital endowment for live trading. The STDAW framework alone requires sophisticated CI/CD infrastructure.

Extremely low reproducibility. The study is a 20-month longitudinal live trading experiment with proprietary brokerage data explicitly withheld. STDAW module is closed-source. The framework is described as open-source at QuantPits.com but the core alignment mechanism (RO-Lock, constraint matrix) is not publicly available. The methodology relies heavily on human-in-the-loop interventions that are difficult to replicate. No controlled baseline comparisons are provided (uses self-evolution as comparison). The 'RL' framing is conceptual rather than algorithmic (no gradient updates). Authors are the sole operators and funders.