ADR-011: Foundry V9 — Oktopus Redesign
Status: Accepted · 2026-04-29
Supersedes: V8 Oktopus Architecture
Context
After 150+ strategies and 8 evolution versions, zero strategies pass the new WF-Gate V8.1. The old gate rewarded "not losing" over "making money":
- +0.001% counted same as +5% — profitable_ratio was binary
- 0-1 trades/window passed — statistical noise treated as edge
- All 4 "champions" demoted — V33 (0.4 trades/window), VolSqueeze (OOS -0.55%), V17 (OOS -0.15%)
Key Lessons
| # | Lesson | Evidence |
|---|---|---|
| 1 | MR works only on DOGE/ADA/AVAX | Consistent 0/6 on BTC/ETH across all MR strategies |
| 2 | Trailing Stop 2-3% never fires | 80-91% signal_exit, 1-5% trailing_stop |
| 3 | Entries too restrictive | 0.4-1.3 trades/window (target: 5+) |
| 4 | Non-MR arms = 0 WF-passed | TREND, MOM, REGIME, 4H all failed |
| 5 | IS-Score overfits 150x | V17 IS=4.88 but OOS=-0.15% |
| 6 | Standard indicators have no edge on 1h-crypto | 150+ combinations, 0 real edge |
Decision
V9 Arms (Redesigned)
| Arm | Name | Target | Entry Style | Exit |
|---|---|---|---|---|
| 1 | MR-ALT | DOGE/ADA/AVAX only | Loose MR (RSI<35, BB without filter) | Signal-Reversal |
| 2 | MR-RELAXED | All 6 | Very loose (RSI<40, BB_Lower only) | Signal-Reversal |
| 3 | TREND-REGIME | All 6 | Only in trending regime (ADX>20 + EMA slope) | Signal-Reversal |
| 4 | SIGNAL-EXIT | All 6 | Paired entry+exit (RSI<30 → RSI>50) | Explicit signal |
| 5 | VOL-BOOSTED | All 6 | Volume + money flow (MFI, CMF, OBV) | Signal-Reversal |
| 6 | VOLATILITY-BREAK | All 6 | Volatility Squeeze + Breakout (BB_Width, ATR, Keltner) | Signal-Reversal |
WF-Gate V8.1 (New Scoring)
| Component | Weight | Range | Floor |
|---|---|---|---|
| OOS Return | 40% | 0..+2% | 0 (negative = fail) |
| Profitable Ratio | 30% | 0..1 | ≥2/3 target assets |
| Trade Floor | 30% | 3..50 trades/window | <3 = fail |
PASS requires ALL of: - robustness ≥ 40 - OOS > 0 on target assets - ≥ 3 trades/asset/window average - ≥ 2/3 target assets profitable (MR) or ≥ 3/6 (non-MR)
Composite Fitness V9
| Component | Weight | Range | Floor |
|---|---|---|---|
| OOS Return | 30% | 0..+2% | 0 (negative = 0) |
| Target-Asset Profitable | 25% | per target group | — |
| Trade Quality | 20% | 3..50 | <3 = 0 |
| WF Robustness | 15% | 0..100 | — |
| Drawdown | 10% | 0..20% | >20% = 0 |
Mutation Guards
- Only mutate HOF members — no random base
- Entry-only mutations — exit is Signal-Reversal (fixed per arm)
- Base must have ≥3 trades/asset/window before mutation
- MR max 40% of HOF — enforced by quota
- Temperature per arm — MR=0.3-0.4, TREND/CROSS=0.5-0.6
Exit Overhaul
- Trailing Stop REMOVED — never fires at 2-3% on 1h
- Signal-Reversal-Exit: explicit exit condition (RSI>50, close>bb_mid, close>ema_20)
- Stop Loss stays (3-4%) and Max Hold stays (18-24 bars)
- Arm 4 (SIGNAL-EXIT): LLM generates both entry AND exit as a pair
IS Pre-Filter
- Strategies with <3 trades/asset/window skip WF entirely
- Saves ~80% compute time on noise strategies
HOF
- Starts EMPTY — all V8 entries failed V8.1 gate
- Per-asset OOS breakdown stored in entries
- Signal-exit vs stop-loss exit ratios tracked
Consequences
Positive: - Eliminates "not losing" false champions - Asset-specific MR targets volatile alts where it works - Signal-Reversal exits match actual exit behavior - Pre-filter saves compute on noise strategies
Negative: - May still find 0 strategies — 1h standard indicators may simply have no edge - Asset-specific targets make comparison harder - Signal-Reversal needs explicit exit conditions in DSL
Neutral: - V8 code preserved as reference - V9 is backward-compatible with backtest engine (signal_exit already dominant)