Systematic Framework & Quantitative Playbook: Replayable Degradable Ai Investment Research Agent

In high-uncertainty markets and rapidly evolving technical environments, mastering 從 Perplexity、Codex 到 MCP 2.0：如何建立可回放、可降級且不會競態覆寫的 AI 投資研究代理 requires moving beyond intuition toward a repeatable, quantitative operating system. Written from the perspective of the Aciemind Digital Clone, this guide translates 不只討論多代理協作，而是聚焦投資研究代理的端到端可靠性：Perplexity 負責資料蒐集、Codex 生成分析、Hermes Agent 執行任務、MCP 串接外部工具，並用狀態機、版本向量、If-Match 或分散式鎖避免研究結論被舊版本覆蓋。加入來源過期、工具失敗、部分完成與人工接管的可回放降級案例，對照記憶中的 post502、case4 與 post171 驗收紀錄，區分參考架構與本站實際部署成果。 into an auditable decision framework with strict risk boundaries.

1. The Core Problem: Why Intuition Fails Under Tail Risk

Whether managing an options selling portfolio (Cash-Secured Puts, Iron Condors, and the Wheel Strategy), architecting edge AI agents on Cloudflare Workers, or executing multi-cycle capital allocation, the most common failure mode is conflating short-term variance with long-term expected value (EV).

1.1 Structural Asymmetries

• Volatility and Liquidity Mismatch: Entering positions without evaluating IV Rank alongside Gamma exposure and Days to Expiration (DTE) exposes the portfolio to convex tail drawdowns.
• Absence of Deterministic Gatekeepers: Relying on discretionary judgment during high-stress market regimes leads to emotional overrides.
• Capital Efficiency Decay: Holding positions into the final 14 DTE exposes capital to accelerating Gamma risk for diminishing Theta returns.

---

2. Four-Stage Closed-Loop Decision Architecture

To make execution deterministic, we formalize the workflow into a four-stage state machine: Regime Sensing → Risk Gatekeeping → Position Sizing → Dynamic Defense & Memory Feedback.

flowchart TD
    A["Stage 1: Volatility & Regime Sensing"] --> B{"Stage 2: Positive EV & Margin Gatekeeper"}
    B -- "Filtered Out (Hold Cash)" --> C["Preserve Buying Power & Reserve"]
    B -- "Passes Risk Threshold" --> D["Stage 3: Automated Staggered Execution"]
    D --> E["Stage 4: Real-Time Greeks & Telemetry Monitoring"]
    E --> F{"Defense Trigger Reached?"}
    F -- "50%~65% Max Profit Captured" --> G["Close Early & Recycle Capital"]
    F -- "Delta > 0.32 or 21 DTE" --> H["Execute Net-Credit Roll Out & Down"]
    G --> I["Commit Outcome to Clone Memory Bank"]
    H --> I

---

3. Quantitative Scenario Stress-Testing

We evaluate performance across three distinct market regimes using a standardized USD 50,000 allocation model:

Dimension	Scenario A: Favorable Convergence	Scenario B: Range-Bound Theta Decay	Scenario C: Tail Volatility Shock
Primary Driver	Directional & IV compression	Steady daily Theta capture	Sharp Vega expansion
Systematic Action	Auto-close at 50% ~ 65% profit target	Roll or close at 21 DTE before Gamma spike	Roll out in time for net credit & widen strikes
Max Drawdown Cap	< 1.5%	< 3.2%	Strictly bounded within 5% risk budget
Long-Term EV Impact	Rapid capital velocity	Core recurring cashflow engine	Preserves principal to sell elevated IV

---

4. Execution Checklist

Before deploying capital or shipping to production, verify all six gatekeeper criteria:

• [ ] 1. Regime Verification: Does current IV Rank / system telemetry justify deployment?
• [ ] 2. Two-Sigma Stress Test: Does buying power remain above 40% under a 2-standard-deviation shock?
• [ ] 3. Automated Profit Taking: Is a GTC limit order placed at 50% ~ 65% of maximum credit?
• [ ] 4. Pre-Computed Roll Plan: Are net-credit roll strikes identified if Delta exceeds 0.30 or DTE reaches 21?
• [ ] 5. Binary Event Filter: Are earnings releases and FOMC announcements accounted for in position sizing?
• [ ] 6. Telemetry & Post-Mortem Logging: Is the thesis recorded in the structured memory bank for self-evolution?