Phoenix Iterate

A structured workflow for AI-assisted quantitative strategy iteration. Instead of randomly tweaking code, this skill enforces a disciplined loop: read history, form hypothesis, write code, validate against past mistakes, backtest progressively, diagnose results, and record lessons.

Dependencies

This skill is an orchestrator — it coordinates two companion skills that must be installed separately:

| Skill | Purpose | Commands provided | |-------|---------|------------------| | backtest-poller | Backtest submission, monitoring, early-stop, results | cli.py submit/status/results/logs | | qc-deep-feature-forensics | Deep feature-level forensic analysis | deep_forensics.py |

Install them first:

clawhub install backtest-poller
clawhub install qc-deep-feature-forensics

All cli.py references below assume backtest-poller is installed as a sibling directory (../backtest-poller/cli.py). Adjust the path if your install location differs.

When to use

• "Let's iterate on the strategy"
• "Start a new strategy version"
• "Run the phoenix iteration loop"
• "What should I try next?"

The Iteration Loop

Step 1: Briefing     ──> Read constitutional memory + history
Step 2: Hypothesis   ──> Pick ONE mutation dimension, form testable hypothesis
Step 3: Code         ──> Write strategy, scan for violations
Step 4: Submit       ──> Progressive validation (smoke -> stress -> medium -> full)
Step 5: Diagnose     ──> Run forensics on results
Step 6: Record       ──> Update constitutional memory with new lessons
         |
         └──────────> Loop back to Step 1

Step 1: Get Decision Context

python3 orchestrator.py briefing

Returns:

• Constitutional memory (all past lessons sorted by severity)
• Available strategy blueprints with historical performance
• Mutation dimensions (what can be changed)
• Recent iteration history (what was tried, what happened)
• Resource constraints and validation windows

Step 2: Form a Trading Hypothesis

Before writing code, answer these questions:

1. What am I changing? — Pick exactly one mutation dimension
2. Why do I think this will be better? — Based on data, not intuition
3. What if I'm wrong? — Worst case scenario, does it violate any iron laws?
4. Will this affect trade frequency? — If trades decrease, there must be a compensating mechanism

Mutation Dimensions:

| Dimension | Description | Examples | |-----------|-------------|---------| | survival_structure | How the strategy survives downturns | QQQ buy-and-hold, cash defense, SMA filter | | position_sizing | How much capital per trade | Kelly %, max trade value, total exposure cap | | selection | What stocks to trade | Premarket volume, RS breakout, HV lottery | | profit_management | What to do with winners | Hold to expiry, take-profit at 200%, trailing stop |

Step 3: Write & Scan Code

1. Create a new strategy directory: strategies/M{N}_{description}/main.py
2. Base it on the latest passing version
3. Run constitutional scan:

python3 orchestrator.py scan --code strategies/M31_test/main.py

If violations are found, fix them before submitting.

Step 4: Progressive Validation

Submit through increasingly demanding test windows. Only promote to the next stage after passing.

Stage 1: Smoke Test  (3 months,  ~15min)  DD < 50%  — Catch obvious bugs
Stage 2: Stress Test (5 months,  ~30min)  DD < 45%  — Survive worst conditions
Stage 3: Medium      (18 months, ~1hr)    DD < 42%  — Bull/bear transitions
Stage 4: Full        (3 years,   ~3hr)    DD < 40%  — Final acceptance

# Check no active backtests  (backtest-poller skill)
python3 ../backtest-poller/cli.py status

# Submit with early-stop protection
python3 ../backtest-poller/cli.py submit \
  --backtest-id <id> \
  --name "MyStrategy_smoke" \
  --max-dd 50

# Monitor progress
python3 ../backtest-poller/cli.py status

# View results when done
python3 ../backtest-poller/cli.py results --name "MyStrategy_smoke" --full

Shortcut: If you only changed profit management (not entry logic), you can skip Smoke/Stress and start from Medium.

Step 5: Diagnose

python3 orchestrator.py diagnose \
  --orders results/MyStrategy_smoke/xxx_orders.csv \
  --result results/MyStrategy_smoke/xxx_result.json

For deeper analysis, run the feature forensics (requires qc-deep-feature-forensics skill):

python3 ../qc-deep-feature-forensics/deep_forensics.py results/MyStrategy_smoke/xxx_orders.csv

Step 6: Record & Decide

python3 orchestrator.py record \
  --name "MyStrategy_v2" \
  --blueprint "baseline" \
  --dimension "position_sizing" \
  --hypothesis "Reduce position size to 2% for better survival" \
  --window "smoke_test" \
  --status "completed" \
  --sharpe 1.5 \
  --drawdown 0.35 \
  --net-profit 0.85

Decision Matrix:

| Outcome | Action | |---------|--------| | Current stage passed | Change dates to next stage, resubmit | | DD exceeded but promising | Analyze monthly cashflow, find bleeding point, targeted fix | | Too few trades | Red flag — loosen entry conditions or shorten cooldown | | Sharpe < 1.0 but DD ok | Acceptable for now, optimize Sharpe later | | Full period passed | Strategy qualified! |

Quick Reference

# This skill (orchestrator)
python3 orchestrator.py briefing          # Decision context
python3 orchestrator.py status             # System status
python3 orchestrator.py list               # Available strategies
python3 orchestrator.py scan --code <path> # Code violation check

# backtest-poller skill
python3 ../backtest-poller/cli.py status                      # Backtest status
python3 ../backtest-poller/cli.py logs --lines 30             # Poller logs
python3 ../backtest-poller/cli.py results --name <name> --full # View results

# qc-deep-feature-forensics skill
python3 ../qc-deep-feature-forensics/deep_forensics.py <orders.csv>  # Deep forensics

Rules

• Only 1 backtest can run at a time (serial execution). Never submit a second backtest while one is running.
• Full-period backtests take 2-3 hours — always use progressive validation to fail fast. Do not jump directly to full-period.
• Always run scan_strategy() before submitting. Constitutional violations must be resolved before backtest submission. Do not bypass the scanner.
• Each iteration must change exactly one mutation dimension. Changing multiple variables at once makes it impossible to attribute success or failure to a specific change.
• Record every iteration result, including failures. Unrecorded iterations will be lost and the same mistake may be repeated.
• The poller daemon handles monitoring automatically after submission. Do not poll the QC API manually — let the daemon manage state.