Day 85: What We Learned - January 21, 2026
Day 85 of 90 | Wednesday, January 21, 2026
Answer Block
Answer Block: 5 days remaining in our journey to build a profitable AI trading system.
5 days remaining in our journey to build a profitable AI trading system.
Today was a wake-up call. Two critical issues surfaced that could have derailed our entire trading operation. Here’s what went wrong and how we’re fixing it.
The Hard Lessons
These are the moments that test us. Critical issues that demanded immediate attention.
Strategy Violation Crisis - Multiple Rogue Workflows
On Jan 21, 2026, the trading system LOST $70.13 due to executing trades that VIOLATE CLAUDE.md strategy mandate. The system bought SPY SHARES and SOFI OPTIONS when it should ONLY execute iron condors
Key takeaway: Portfolio: $5,028.
CTO Failure - Stale Data Led to Misinformation
CTO (Claude) gave CEO incorrect P/L information multiple times:
Key takeaway: Claimed $0.
Position Imbalance Crisis - Orphan Long Puts
Portfolio lost $329.42 (-6.59%) due to position imbalance:
Key takeaway: The orphan longs are decaying and losing money without corresponding short premium to offset.
SOFI Position Blocked All Trading - Buying Power Crisis
- SOFI260213P00032000 (short put) was open with -$685 market value
Key takeaway: 1. Triggered close-non-spy-positions.yml workflow
Position Limit - Count Contracts Not Symbols
The position limit check was counting UNIQUE SYMBOLS instead of TOTAL CONTRACTS:
Key takeaway: 3. Log details: Show exact positions when limit reached
Partial Iron Condor Auto-Close
Iron condors were being placed with only PUT legs filling. CALL legs were failing silently, leaving dangerous directional positions:
Important Discoveries
Not emergencies, but insights that will shape how we trade going forward.
CALL Leg Pricing Fix - Aggressive Fallbacks
Iron condors were placing PUT legs successfully but CALL legs were failing:
Iron Condor Optimization Research - 86% Win Rate Strategy
- Options Trading IQ: Iron Condor Success Rate
- [Project Finance: Iron Condor Management (71,417 trades)](https://www.projectfinance.com/iron
Quick Wins & Refinements
- Day 2 Crisis - Position Imbalance and Missing CALL Legs - Two consecutive days of trading crises:…
Today’s Numbers
| What | Count |
|---|---|
| Lessons Learned | 9 |
| Critical Issues | 6 |
| High Priority | 2 |
| Improvements | 1 |
Tech Stack Behind the Lessons
Every lesson we learn is captured, analyzed, and stored by our AI infrastructure:
Detected"] --> CLAUDE["Claude Opus
(Analysis)"] CLAUDE --> RAG["LanceDB RAG
(Storage)"] RAG --> BLOG["GitHub Pages
(Publishing)"] BLOG --> DEVTO["Dev.to
(Distribution)"] end
How We Learn Autonomously
| Component | Role in Learning |
|---|---|
| Claude Opus 4.5 | Analyzes errors, extracts insights, determines severity |
| LanceDB RAG | Stores lessons with 768D embeddings for semantic search |
| Gemini 2.0 Flash | Retrieves relevant past lessons before new trades |
| OpenRouter (DeepSeek) | Cost-effective sentiment analysis and research |
Why This Matters
- No Lesson Lost: Every insight persists in our RAG corpus
- Contextual Recall: Before each trade, we query similar past situations
- Continuous Improvement: 200+ lessons shape every decision
- Transparent Journey: All learnings published publicly
The Journey So Far
We’re building an autonomous AI trading system that learns from every mistake. This isn’t about getting rich quick - it’s about building a system that can consistently generate income through disciplined options trading.
Our approach:
- Paper trade for 90 days to validate the strategy
- Document every lesson, every failure, every win
- Use AI (Claude) as CTO to automate and improve
- Follow Phil Town’s Rule #1: Don’t lose money
Want to follow along? Check out the full project on GitHub.
FAQ
What did we learn today?
9 lessons captured (6 critical, 2 high). Today was a wake-up call. Two critical issues surfaced that could have derailed our entire trading operation. Here’s what went wrong and how we’re fixing it.
How do you keep these lessons from getting lost?
We index every lesson into a RAG corpus and query it before new trades and major engineering changes.
Where is the canonical version of this post?
This post’s canonical URL is https://igorganapolsky.github.io/trading/2026/01/21/lessons-learned/.
Day 85/90 complete. 5 to go.