diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
index 61152ae..2ad2872 100644
--- a/.github/copilot-instructions.md
+++ b/.github/copilot-instructions.md
@@ -167,6 +167,11 @@ Without this, the SDK returns signatures for transactions that never execute, ca
 
 **Important fields:**
 - `signalQualityScore` (Int?) - 0-100 score for data-driven optimization
+- `signalQualityVersion` (String?) - Tracks which scoring logic was used ('v1', 'v2', 'v3')
+  - v1: Original logic (price position < 5% threshold)
+  - v2: Added volume compensation for low ADX (2025-11-07)
+  - v3: CURRENT - Stricter breakdown requirements: positions < 15% require (ADX > 18 AND volume > 1.2x) OR (RSI < 35 for shorts / RSI > 60 for longs)
+  - All new trades tagged with current version for comparative analysis
 - `maxFavorableExcursion` / `maxAdverseExcursion` - Track best/worst P&L during trade lifetime
 - `maxFavorablePrice` / `maxAdversePrice` - Track prices at MFE/MAE points
 - `configSnapshot` (Json) - Stores Position Manager state for crash recovery
@@ -219,6 +224,7 @@ const driftSymbol = normalizeTradingViewSymbol(body.symbol)
 - `/api/trading/positions` - Query open positions from Drift
 - `/api/settings` - Get/update config (writes to .env file, **includes per-symbol settings**)
 - `/api/analytics/last-trade` - Fetch most recent trade details for dashboard (includes quality score)
+- `/api/analytics/version-comparison` - Compare performance across signal quality logic versions (v1/v2/v3)
 - `/api/restart` - Create restart flag for watch-restart.sh script
 
 ## Critical Workflows
@@ -412,6 +418,13 @@ docker exec trading-bot-postgres psql -U postgres -d trading_bot_v4 -c "\dt"
     - Bot's quality scoring provides second-layer filtering with context-aware metrics
     - Two-stage filtering (TradingView + bot) prevents both overtrading and missing valid signals
 
+18. **Prisma Decimal type handling:** Raw SQL queries return Prisma `Decimal` objects, not plain numbers:
+    - Use `any` type for numeric fields in `$queryRaw` results: `total_pnl: any`
+    - Convert with `Number()` before returning to frontend: `totalPnL: Number(stat.total_pnl) || 0`
+    - Frontend uses `.toFixed()` which doesn't exist on Decimal objects
+    - Applies to all aggregations: SUM(), AVG(), ROUND() - all return Decimal types
+    - Example: `/api/analytics/version-comparison` converts all numeric fields
+
 ## File Conventions
 
 - **API routes:** `app/api/[feature]/[action]/route.ts` (Next.js 15 App Router)
@@ -477,6 +490,12 @@ See `POSITION_SCALING_ROADMAP.md` for planned optimizations:
 
 **Data-driven approach:** Each phase requires validation through SQL analysis before implementation. No premature optimization.
 
+**Signal Quality Version Tracking:** Database tracks `signalQualityVersion` field to compare algorithm performance:
+- Analytics dashboard shows version comparison: trades, win rate, P&L, extreme position stats
+- Focus on extreme positions (< 15% range) - v3 aims to reduce losses from weak ADX entries
+- SQL queries in `docs/analysis/SIGNAL_QUALITY_VERSION_ANALYSIS.sql` for deep-dive analysis
+- Need 20+ v3 trades before meaningful comparison vs v1/v2 data
+
 ## Integration Points
 
 - **n8n:** Expects exact response format from `/api/trading/execute` (see n8n-complete-workflow.json)