diff --git a/.dockerignore b/.dockerignore
new file mode 100644
index 0000000..8ecc131
--- /dev/null
+++ b/.dockerignore
@@ -0,0 +1,11 @@
+.git/
+__pycache__/
+*.pyc
+.env
+logs/
+models/
+.claude/
+venv/
+data/
+fx-quant/
+*.pdf
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..2e5074c
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,19 @@
+FROM python:3.11-slim
+
+# LightGBM needs libgomp
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgomp1 && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY requirements.docker.txt .
+RUN pip install --no-cache-dir -r requirements.docker.txt
+
+COPY config/system.yaml config/system.yaml
+COPY src/ src/
+COPY sql/ sql/
+
+RUN mkdir -p logs models
+
+ENV PYTHONUNBUFFERED=1
+CMD ["python", "src/order_executor.py"]
diff --git a/config/system.yaml b/config/system.yaml
index 5c076f4..e01f626 100644
--- a/config/system.yaml
+++ b/config/system.yaml
@@ -26,10 +26,16 @@ features:
   volatility_window: 20
 
 ai:
-  model: "local-ensemble" # example: 'openai', 'local-llm', 'ensemble'
+  model: "local-ensemble"
   confidence_threshold: 0.85
   retriever_enabled: true
   retriever_source: "supabase"
+  ensemble_models: ["logistic_regression", "random_forest", "gradient_boosting"]
+  backtest_validation_window: 50
+  sanity_checks:
+    rsi_overbought: 80
+    rsi_oversold: 20
+    volatility_multiplier: 3.0
 
 strategy:
   rule: "sma_cross" # human-readable name of the rule
@@ -43,6 +49,7 @@ execution:
   paper_mode: true
   canary_size_pct: 0.01
   max_positions: 3
+  interval_seconds: 60
 
 supabase:
   url: "https://<your>.supabase.co"
diff --git a/docker-compose.yml b/docker-compose.yml
new file mode 100644
index 0000000..830ef65
--- /dev/null
+++ b/docker-compose.yml
@@ -0,0 +1,11 @@
+services:
+  fx-quant:
+    build: .
+    container_name: fx-quant
+    env_file:
+      - config/.env
+    volumes:
+      - ./config/system.yaml:/app/config/system.yaml:ro
+      - ./logs:/app/logs
+      - ./models:/app/models
+    restart: unless-stopped
diff --git a/requirements.docker.txt b/requirements.docker.txt
new file mode 100644
index 0000000..bc463b8
--- /dev/null
+++ b/requirements.docker.txt
@@ -0,0 +1,13 @@
+pandas>=2.0
+numpy>=1.24
+scipy>=1.10
+scikit-learn==1.8.0
+lightgbm>=4.0
+matplotlib>=3.7
+plotly>=5.0
+pyarrow>=14.0
+oandapyV20==0.7.2
+requests>=2.28
+supabase>=2.0
+python-dotenv>=1.0
+PyYAML>=6.0
diff --git a/src/ai_wrapper.py b/src/ai_wrapper.py
new file mode 100644
index 0000000..a652360
--- /dev/null
+++ b/src/ai_wrapper.py
@@ -0,0 +1,492 @@
+# src/ai_wrapper.py
+"""
+AI Decision Wrapper for fx-quant.
+Validates trading signals using a local ML ensemble (scikit-learn + LightGBM)
+before allowing order execution. Low-confidence or sanity-failed signals
+are rejected and logged for human review.
+"""
+
+import csv
+import warnings
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from sklearn.linear_model import LogisticRegression
+from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
+
+from config_loader import load_config, get_project_root
+from backtester import run_backtest, generate_signals
+
+# Try to use LightGBM if available, otherwise fall back to sklearn GBM
+try:
+    from lightgbm import LGBMClassifier
+    HAS_LGBM = True
+except ImportError:
+    HAS_LGBM = False
+
+
+# ---------------------------------------------------------------------------
+# Feature engineering
+# ---------------------------------------------------------------------------
+
+FEATURE_COLS = ["sma_3", "sma_20", "rsi_14", "atr_14", "vol_20", "ret", "close"]
+
+
+def build_feature_vector(df, idx=-1):
+    """
+    Extract model input features from the signal DataFrame at a given bar index.
+
+    Features are normalized by dividing price-scale columns by close price.
+    Returns a 1D numpy array ready for model prediction.
+    """
+    row = df.iloc[idx]
+    close = row["close"]
+    if close == 0:
+        close = 1e-10  # avoid division by zero
+
+    features = []
+    for col in FEATURE_COLS:
+        val = row.get(col, 0.0)
+        if pd.isna(val):
+            val = 0.0
+        # Normalize price-scale features relative to close
+        if col in ("sma_3", "sma_20", "atr_14"):
+            val = val / close
+        features.append(float(val))
+
+    return pd.DataFrame([features], columns=FEATURE_COLS)
+
+
+# ---------------------------------------------------------------------------
+# Context builder
+# ---------------------------------------------------------------------------
+
+def build_context(df, instrument, supabase_client=None, cfg=None):
+    """
+    Assemble retrieval context: last 50 bars of features, recent backtest trades.
+    Used for rationale generation in decision logging.
+
+    Returns dict with context summary.
+    """
+    cfg = cfg or {}
+    ai_cfg = cfg.get("ai", {})
+    context = {
+        "instrument": instrument,
+        "bars_available": len(df),
+        "last_close": float(df["close"].iloc[-1]) if len(df) > 0 else None,
+    }
+
+    # Recent feature summary (last 50 bars)
+    tail = df.tail(50)
+    for col in FEATURE_COLS:
+        if col in tail.columns:
+            context[f"{col}_mean"] = float(tail[col].mean()) if not tail[col].isna().all() else None
+
+    # Optional: pull recent rows from Supabase for retrieval context
+    if ai_cfg.get("retriever_enabled") and supabase_client is not None:
+        try:
+            table = cfg.get("supabase", {}).get("table", "fx_candles")
+            resp = (
+                supabase_client.table(table)
+                .select("time,close,ret")
+                .eq("instrument", instrument)
+                .order("time", desc=True)
+                .limit(10)
+                .execute()
+            )
+            recent = resp.data or []
+            context["recent_supabase_rows"] = len(recent)
+        except Exception as e:
+            context["recent_supabase_rows"] = 0
+            context["retriever_error"] = str(e)
+
+    return context
+
+
+# ---------------------------------------------------------------------------
+# Ensemble training
+# ---------------------------------------------------------------------------
+
+def _build_feature_matrix(df):
+    """Build feature matrix from DataFrame. Returns (X, feature_names)."""
+    close = df["close"].replace(0, 1e-10)
+    X = pd.DataFrame(index=df.index)
+    for col in FEATURE_COLS:
+        if col not in df.columns:
+            X[col] = 0.0
+        elif col in ("sma_3", "sma_20", "atr_14"):
+            X[col] = df[col] / close
+        else:
+            X[col] = df[col]
+    X = X.fillna(0.0)
+    return X, list(X.columns)
+
+
+def train_ensemble(df, strategy_cfg):
+    """
+    Train 3 models on historical data:
+      - LogisticRegression
+      - RandomForestClassifier
+      - GradientBoostingClassifier (or LGBMClassifier if available)
+
+    Target: whether the signal led to a profitable next bar
+    (next-bar return > 0 when signal == 1).
+
+    Uses walk-forward split: train on first 70%, validate on last 30%.
+    Returns (list of fitted model objects, validation metrics dict).
+    """
+    # Need signal column
+    if "signal" not in df.columns:
+        df = generate_signals(df, strategy_cfg)
+
+    # Build target: next-bar return is positive AND we have a long signal
+    next_ret = df["ret"].shift(-1)
+    target = ((next_ret > 0) & (df["signal"] == 1)).astype(int)
+
+    # Drop last row (no next-bar return) and any NaN rows
+    valid_mask = next_ret.notna()
+    df_valid = df[valid_mask]
+    target = target[valid_mask]
+
+    X, feature_names = _build_feature_matrix(df_valid)
+    y = target.values
+
+    if len(X) < 50:
+        print(f"  AI: Not enough data to train ensemble ({len(X)} rows). Skipping.")
+        return [], {"error": "insufficient_data", "rows": len(X)}
+
+    # Walk-forward split: 70/30
+    split_idx = int(len(X) * 0.7)
+    X_train, X_val = X[:split_idx], X[split_idx:]
+    y_train, y_val = y[:split_idx], y[split_idx:]
+
+    # Train models
+    models = []
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+
+        lr = LogisticRegression(max_iter=500, random_state=42)
+        lr.fit(X_train, y_train)
+        models.append(("logistic_regression", lr))
+
+        rf = RandomForestClassifier(n_estimators=100, max_depth=5, random_state=42)
+        rf.fit(X_train, y_train)
+        models.append(("random_forest", rf))
+
+        if HAS_LGBM:
+            gbm = LGBMClassifier(n_estimators=100, max_depth=5, random_state=42, verbose=-1)
+        else:
+            gbm = GradientBoostingClassifier(n_estimators=100, max_depth=5, random_state=42)
+        gbm.fit(X_train, y_train)
+        models.append(("gradient_boosting", gbm))
+
+    # Validation metrics
+    val_scores = {}
+    for name, model in models:
+        score = model.score(X_val, y_val)
+        val_scores[name] = round(score, 4)
+
+    print(f"  AI: Ensemble trained — validation scores: {val_scores}")
+    return models, val_scores
+
+
+# ---------------------------------------------------------------------------
+# Ensemble prediction
+# ---------------------------------------------------------------------------
+
+def ensemble_predict(models, feature_vector):
+    """
+    Each model predicts probability of profitable outcome.
+
+    Returns (confidence, predictions, agreement):
+      - confidence: mean probability across models
+      - predictions: list of per-model probabilities
+      - agreement: all models agree on direction (all > 0.5 or all < 0.5)
+    """
+    if not models:
+        return 0.0, [], False
+
+    predictions = []
+    for name, model in models:
+        proba = model.predict_proba(feature_vector)
+        # probability of class 1 (profitable)
+        p = float(proba[0][1]) if proba.shape[1] > 1 else float(proba[0][0])
+        predictions.append(p)
+
+    confidence = float(np.mean(predictions))
+    # Agreement: all above 0.5 or all below 0.5
+    all_bullish = all(p > 0.5 for p in predictions)
+    all_bearish = all(p <= 0.5 for p in predictions)
+    agreement = all_bullish or all_bearish
+
+    return confidence, predictions, agreement
+
+
+# ---------------------------------------------------------------------------
+# Sanity checks
+# ---------------------------------------------------------------------------
+
+def sanity_checks(df, instrument, cfg):
+    """
+    Run deterministic sanity checks on the latest bar:
+      - RSI not in extreme territory against the signal
+      - Volatility not excessively high (ATR > 3x rolling mean)
+      - Price not gapping (|ret| > 5x vol)
+
+    Returns (passed, reasons) where reasons lists any failed checks.
+    """
+    ai_cfg = cfg.get("ai", {})
+    sc_cfg = ai_cfg.get("sanity_checks", {})
+    rsi_ob = sc_cfg.get("rsi_overbought", 80)
+    rsi_os = sc_cfg.get("rsi_oversold", 20)
+    vol_mult = sc_cfg.get("volatility_multiplier", 3.0)
+
+    reasons = []
+    latest = df.iloc[-1]
+    signal = int(latest.get("signal", 0))
+
+    # RSI check
+    rsi = latest.get("rsi_14", 50.0)
+    if pd.notna(rsi):
+        if signal == 1 and rsi > rsi_ob:
+            reasons.append(f"RSI={rsi:.1f} > {rsi_ob} (overbought) conflicts with BUY signal")
+        elif signal == 0 and rsi < rsi_os:
+            reasons.append(f"RSI={rsi:.1f} < {rsi_os} (oversold) conflicts with FLAT signal")
+
+    # ATR volatility check
+    atr = latest.get("atr_14", None)
+    if atr is not None and pd.notna(atr) and "atr_14" in df.columns:
+        atr_mean = df["atr_14"].rolling(50).mean().iloc[-1]
+        if pd.notna(atr_mean) and atr_mean > 0 and atr > vol_mult * atr_mean:
+            reasons.append(f"ATR={atr:.6f} > {vol_mult}x rolling mean ({atr_mean:.6f})")
+
+    # Gap check: |ret| > 5x vol_20
+    ret = latest.get("ret", 0.0)
+    vol = latest.get("vol_20", None)
+    if vol is not None and pd.notna(vol) and vol > 0 and pd.notna(ret):
+        if abs(ret) > 5.0 * vol:
+            reasons.append(f"|ret|={abs(ret):.6f} > 5x vol_20 ({vol:.6f}), possible gap")
+
+    passed = len(reasons) == 0
+    return passed, reasons
+
+
+# ---------------------------------------------------------------------------
+# Backtest recent window
+# ---------------------------------------------------------------------------
+
+def backtest_recent_window(df, strategy_cfg, window=50):
+    """
+    Run backtester on the last N bars to check if the strategy
+    is profitable in the recent window.
+
+    Returns (profitable, metrics).
+    """
+    if len(df) < window:
+        window = len(df)
+
+    if window < 10:
+        return False, {"error": "insufficient_data", "bars": window}
+
+    recent_df = df.tail(window).copy()
+
+    try:
+        results = run_backtest(recent_df, strategy_cfg)
+        metrics = results["metrics"]
+        profitable = metrics["total_return_pct"] > 0
+        return profitable, metrics
+    except Exception as e:
+        return False, {"error": str(e)}
+
+
+# ---------------------------------------------------------------------------
+# Signal validation (main entry point)
+# ---------------------------------------------------------------------------
+
+def validate_signal(instrument, signal, df, cfg, models=None):
+    """
+    Main entry point — orchestrates all AI validation checks:
+      1. Build feature vector -> ensemble predict -> confidence + agreement
+      2. Run sanity checks
+      3. Run backtest on recent window
+      4. Combine into final decision
+
+    If confidence < threshold OR sanity fails OR ensemble disagrees -> REJECT.
+
+    Returns dict with: approved, confidence, rationale, checks.
+    """
+    ai_cfg = cfg.get("ai", {})
+    threshold = ai_cfg.get("confidence_threshold", 0.85)
+    bt_window = ai_cfg.get("backtest_validation_window", 50)
+    strategy_cfg = cfg.get("strategy", {})
+
+    checks = {}
+    rationale_parts = []
+
+    # 1. Ensemble prediction
+    if models:
+        fv = build_feature_vector(df)
+        confidence, predictions, agreement = ensemble_predict(models, fv)
+        checks["ensemble_confidence"] = round(confidence, 4)
+        checks["ensemble_predictions"] = [round(p, 4) for p in predictions]
+        checks["ensemble_agreement"] = agreement
+    else:
+        confidence = 0.0
+        agreement = False
+        checks["ensemble_confidence"] = 0.0
+        checks["ensemble_agreement"] = False
+        rationale_parts.append("No ensemble models available")
+
+    # 2. Sanity checks
+    sanity_passed, sanity_reasons = sanity_checks(df, instrument, cfg)
+    checks["sanity_passed"] = sanity_passed
+    checks["sanity_reasons"] = sanity_reasons
+    if not sanity_passed:
+        rationale_parts.append(f"Sanity failed: {'; '.join(sanity_reasons)}")
+
+    # 3. Backtest recent window
+    bt_profitable, bt_metrics = backtest_recent_window(df, strategy_cfg, window=bt_window)
+    checks["backtest_profitable"] = bt_profitable
+    checks["backtest_metrics"] = bt_metrics
+    if not bt_profitable:
+        rationale_parts.append(f"Recent backtest unprofitable: {bt_metrics.get('total_return_pct', 'N/A')}%")
+
+    # 4. Final decision
+    approved = True
+    if confidence < threshold:
+        approved = False
+        rationale_parts.append(f"Confidence {confidence:.4f} < threshold {threshold}")
+    if not agreement and models:
+        approved = False
+        rationale_parts.append("Ensemble models disagree on direction")
+    if not sanity_passed:
+        approved = False
+
+    rationale = "; ".join(rationale_parts) if rationale_parts else "All checks passed"
+
+    decision = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "instrument": instrument,
+        "signal": signal,
+        "confidence": round(confidence, 4),
+        "approved": approved,
+        "ensemble_agreement": agreement if models else False,
+        "sanity_passed": sanity_passed,
+        "backtest_profitable": bt_profitable,
+        "rationale": rationale,
+        "checks": checks,
+    }
+
+    return decision
+
+
+# ---------------------------------------------------------------------------
+# Decision logging
+# ---------------------------------------------------------------------------
+
+def log_ai_decision(decision_details):
+    """
+    Append AI decision to logs/ai_decisions.csv.
+    Creates the file with headers if it doesn't exist.
+    """
+    root = get_project_root()
+    logs_dir = root / "logs"
+    logs_dir.mkdir(exist_ok=True)
+    csv_path = logs_dir / "ai_decisions.csv"
+
+    columns = [
+        "timestamp", "instrument", "signal", "confidence", "approved",
+        "ensemble_agreement", "sanity_passed", "backtest_profitable", "rationale",
+    ]
+
+    row = {col: decision_details.get(col, "") for col in columns}
+
+    file_exists = csv_path.exists()
+    with open(csv_path, "a", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=columns)
+        if not file_exists:
+            writer.writeheader()
+        writer.writerow(row)
+
+
+# ---------------------------------------------------------------------------
+# Standalone test
+# ---------------------------------------------------------------------------
+
+def main():
+    """
+    Standalone test: train ensemble on Supabase data, run validation,
+    and print results.
+    """
+    import os
+    from supabase import create_client
+    from backtester import fetch_candles_from_supabase
+
+    cfg = load_config()
+
+    print("=" * 60)
+    print("fx-quant AI Decision Wrapper — Standalone Test")
+    print("=" * 60)
+
+    # Supabase client
+    supabase_url = os.getenv("SUPABASE_URL")
+    supabase_key = os.getenv("SUPABASE_KEY")
+    if not supabase_url or not supabase_key:
+        raise SystemExit("Missing SUPABASE_URL or SUPABASE_KEY in config/.env")
+    sb = create_client(supabase_url, supabase_key)
+    table = cfg.get("supabase", {}).get("table", "fx_candles")
+
+    strategy_cfg = cfg["strategy"]
+    instruments = cfg["brokers"][0]["instruments"]
+    granularity = cfg["data"]["candle_granularities"][0]
+
+    for instrument in instruments:
+        print(f"\n--- {instrument} ---")
+
+        df = fetch_candles_from_supabase(instrument, granularity, sb, table)
+        if df.empty:
+            print(f"  No data for {instrument}. Skipping.")
+            continue
+
+        df = generate_signals(df, strategy_cfg)
+        if df.empty:
+            print(f"  No valid rows after signal generation. Skipping.")
+            continue
+
+        # Train ensemble
+        models, val_metrics = train_ensemble(df, strategy_cfg)
+        print(f"  Validation metrics: {val_metrics}")
+
+        if not models:
+            print("  Skipping validation — no models trained.")
+            continue
+
+        # Build context
+        context = build_context(df, instrument, sb, cfg)
+        print(f"  Context: {context}")
+
+        # Validate latest signal
+        latest_signal = int(df["signal"].iloc[-1])
+        decision = validate_signal(instrument, latest_signal, df, cfg, models=models)
+
+        print(f"  Signal: {'LONG' if latest_signal == 1 else 'FLAT'}")
+        print(f"  Approved: {decision['approved']}")
+        print(f"  Confidence: {decision['confidence']:.4f}")
+        print(f"  Agreement: {decision['ensemble_agreement']}")
+        print(f"  Sanity passed: {decision['sanity_passed']}")
+        print(f"  Backtest profitable: {decision['backtest_profitable']}")
+        print(f"  Rationale: {decision['rationale']}")
+
+        # Log decision
+        log_ai_decision(decision)
+        print(f"  Decision logged to logs/ai_decisions.csv")
+
+    print("\n" + "=" * 60)
+    print("Standalone test complete.")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/order_executor.py b/src/order_executor.py
index 1f5fa48..a3ee3e2 100644
--- a/src/order_executor.py
+++ b/src/order_executor.py
@@ -6,7 +6,9 @@ Includes kill switch, position limits, and full order logging.
 """
 
 import os
+import sys
 import csv
+import time
 from datetime import datetime, timezone
 from pathlib import Path
 
@@ -16,6 +18,7 @@ from supabase import create_client
 
 from config_loader import load_config, get_project_root
 from backtester import fetch_candles_from_supabase, generate_signals
+from ai_wrapper import train_ensemble, validate_signal, log_ai_decision
 
 
 # ---------------------------------------------------------------------------
@@ -277,12 +280,15 @@ def place_order(instrument, units, side, cfg, price=None):
 # Signal execution
 # ---------------------------------------------------------------------------
 
-def execute_signals(signals_df, cfg):
+def execute_signals(signals_df, cfg, ai_models=None):
     """
     Takes a DataFrame with signal column (from generate_signals).
     Reads the latest signal per instrument, compares to current positions,
     and places orders for needed changes.
 
+    If ai_models are provided, validates signals through the AI decision
+    wrapper before placing orders.
+
     Returns list of order results.
     """
     paper_mode = cfg.get("execution", {}).get("paper_mode", True)
@@ -327,6 +333,14 @@ def execute_signals(signals_df, cfg):
 
     # Signal=1 means go long, signal=0 means go flat
     if signal == 1 and current_position == 0:
+        # AI validation before placing BUY order
+        if ai_models:
+            ai_decision = validate_signal(instrument, signal, signals_df, cfg, models=ai_models)
+            log_ai_decision(ai_decision)
+            if not ai_decision["approved"]:
+                print(f"  AI REJECTED: confidence={ai_decision['confidence']:.2f}, {ai_decision['rationale']}")
+                return results
+
         units = compute_units(balance, instrument, "BUY", cfg)
         result = place_order(instrument, units, "BUY", cfg, price=close_price)
         results.append(result)
@@ -418,8 +432,17 @@ def main():
         latest_signal = "LONG" if df["signal"].iloc[-1] == 1 else "FLAT"
         print(f"  Latest signal: {latest_signal} (close={df['close'].iloc[-1]:.5f})")
 
-        # Execute
-        results = execute_signals(df, cfg)
+        # Train AI ensemble for this instrument
+        ai_models = None
+        ai_cfg = cfg.get("ai", {})
+        if ai_cfg.get("model") == "local-ensemble":
+            print(f"  Training AI ensemble for {instrument}...")
+            models_list, val_metrics = train_ensemble(df, strategy_cfg)
+            if models_list:
+                ai_models = models_list
+
+        # Execute with AI validation
+        results = execute_signals(df, cfg, ai_models=ai_models)
         all_results.extend(results)
         print()
 
@@ -432,4 +455,19 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    cfg = load_config()
+    interval = cfg.get("execution", {}).get("interval_seconds", 60)
+
+    # One-shot mode: pass --once to run a single iteration and exit
+    if "--once" in sys.argv:
+        main()
+    else:
+        print(f"Running on {interval}s loop. Press Ctrl+C to stop.\n")
+        while True:
+            try:
+                main()
+                print(f"\nSleeping {interval}s until next run...\n")
+                time.sleep(interval)
+            except KeyboardInterrupt:
+                print("\nShutting down gracefully.")
+                break