diff --git a/.gitignore b/.gitignore
index 74e5757..2b0f920 100644
--- a/.gitignore
+++ b/.gitignore
@@ -17,6 +17,10 @@ database/**/*.db
 # Local demo snapshots (large)
 data/processed/
 data/demos/
+data/sequences/
+
+# Reports and Artifacts
+reports/
 
 # Local downloads / raw captures
 output_arena/
diff --git a/PROJECT_REVIEW.md b/PROJECT_REVIEW.md
new file mode 100644
index 0000000..92ba7be
--- /dev/null
+++ b/PROJECT_REVIEW.md
@@ -0,0 +1,64 @@
+# Clutch-IQ 项目评测报告
+
+## 1. 项目概览 (Overview)
+**Clutch-IQ** 是一个针对 CS2 (Counter-Strike 2) 的实时胜率预测系统。项目实现了从原始 Demo 解析、特征提取、模型训练到实时推理的全链路工程。
+
+*   **完成度**: A- (核心功能闭环，MVP 阶段完成)
+*   **技术栈**: Python, Pandas, XGBoost, Flask, Streamlit, Demoparser2
+*   **应用场景**: 战术分析、直播推流、实时辅助
+
+---
+
+## 2. 深度评测 (Detailed Review)
+
+### ✅ 亮点 (Strengths)
+
+#### 1. 优秀的工程化架构
+项目没有将所有代码堆在一个文件里，而是采用了清晰的分层架构：
+*   **ETL 层 (`src/etl/`)**: 实现了流式处理 (`auto_pipeline.py`)，巧妙解决了海量 Demo 的存储痛点，这在个人项目中是非常亮眼的工程实践。
+*   **特征层 (`src/features/`)**: 将特征定义抽离为 `definitions.py`，确保了训练和推理使用同一套标准，有效防止了**训练-推理偏差 (Training-Serving Skew)**。
+*   **模型层 (`src/training/`)**: 实现了训练与验证的解耦 (`train.py` vs `evaluate.py`)，并且严格执行了 Match-Level Splitting，避免了时序数据的泄露问题，体现了扎实的机器学习素养。
+
+#### 2. 扎实的特征工程
+不仅使用了基础的 K/D 数据，还引入了计算几何概念：
+*   **空间特征**: `t_area` (凸包面积) 和 `pincer_index` (夹击指数) 是非常高级的特征，能够捕捉“枪法”之外的“战术”维度。
+*   **经济特征**: 细分了装备价值和现金流，能准确判断 Eco 局与长枪局的优劣势。
+
+#### 3. 完整的文档体系
+项目包含 `AI_FULL_STACK_GUIDE.md` 和 `PROJECT_DEEP_DIVE.md`，不仅有代码，还有思考过程和理论支撑。这对后续维护和团队协作至关重要。
+
+### ⚠️ 改进空间 (Areas for Improvement)
+
+#### 1. 测试覆盖率 (Test Coverage)
+*   **现状**: 虽然有 `evaluate.py` 验证模型效果，但缺乏针对各个函数的**单元测试 (Unit Tests)**。
+*   **风险**: 如果修改了 `spatial.py` 里的凸包算法，可能会悄悄破坏特征计算逻辑，而直到模型准确率下降才能发现。
+*   **建议**: 引入 `pytest`，为特征计算函数编写测试用例。
+
+#### 2. 配置管理 (Configuration Management)
+*   **现状**: 部分路径和参数（如文件路径、模型参数）可能硬编码在代码中。
+*   **建议**: 引入 `config.yaml` 或 `.env` 管理所有可变参数，使项目更容易在不同机器上部署。
+
+#### 3. 异常处理与日志 (Robustness)
+*   **现状**: 虽然有基础的 logging，但在高并发场景下（如 GSI 频繁推送），`inference/app.py` 的健壮性还需加强。
+*   **建议**: 增加请求队列机制，防止瞬间流量冲垮推理服务。
+
+---
+
+## 3. 综合评分 (Scoring)
+
+| 维度 | 评分 (1-10) | 评价 |
+| :--- | :---: | :--- |
+| **架构设计** | **9.0** | 模块清晰，流式处理是加分项。 |
+| **代码质量** | **8.5** | 风格统一，可读性强，函数封装合理。 |
+| **算法深度** | **8.0** | XGBoost 选型准确，特征有新意，仍有提升空间（如时序模型）。 |
+| **完成度** | **8.5** | 核心闭环已跑通，文档齐全。 |
+| **创新性** | **7.5** | 空间特征的应用是亮点，但整体属于经典 ML 范式。 |
+
+### 🏆 总评: 优秀 (Excellent)
+
+**Clutch-IQ** 是一个具备**生产级潜质**的个人项目。它超越了普通的“Demo 代码”，展现了完整的全栈 AI 工程思维。特别是对存储限制的优化和防数据泄露的处理，显示了开发者对实际工程问题的深刻理解。
+
+**下一步建议**:
+1.  **容器化**: 编写 `Dockerfile`，一键部署环境。
+2.  **可视化增强**: 优化 Dashboard，增加特征重要性解释图表（SHAP plots）。
+3.  **实战接入**: 完成 GSI 配置，真正在游戏中跑起来。
diff --git a/models/clutch_attention_lstm_v1.pth b/models/clutch_attention_lstm_v1.pth
new file mode 100644
index 0000000..84ba57a
Binary files /dev/null and b/models/clutch_attention_lstm_v1.pth differ
diff --git a/models/clutch_lstm_v1.pth b/models/clutch_lstm_v1.pth
new file mode 100644
index 0000000..4717e51
Binary files /dev/null and b/models/clutch_lstm_v1.pth differ
diff --git a/src/analysis/ensemble_analysis.py b/src/analysis/ensemble_analysis.py
new file mode 100644
index 0000000..11d7e24
--- /dev/null
+++ b/src/analysis/ensemble_analysis.py
@@ -0,0 +1,90 @@
+import os
+import sys
+import pandas as pd
+import numpy as np
+import torch
+import xgboost as xgb
+from sklearn.metrics import accuracy_score, log_loss
+import matplotlib.pyplot as plt
+import seaborn as sns
+
+# Add project root to path
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+
+from src.training.models import ClutchAttentionLSTM
+from src.features.definitions import FEATURE_COLUMNS
+from src.inference.stacking_ensemble import StackingEnsemble
+
+# Configuration
+XGB_MODEL_PATH = "models/clutch_model_v1.json"
+LSTM_MODEL_PATH = "models/clutch_attention_lstm_v1.pth"
+TEST_DATA_PATH = "data/processed/test_set.parquet"
+
+def analyze_ensemble():
+    if not os.path.exists(TEST_DATA_PATH):
+        print("Test data not found.")
+        return
+
+    print(f"Loading data from {TEST_DATA_PATH}...")
+    df = pd.read_parquet(TEST_DATA_PATH)
+    y = df['round_winner'].values
+
+    # Initialize Ensemble (to reuse get_base_predictions)
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    ensemble = StackingEnsemble(XGB_MODEL_PATH, LSTM_MODEL_PATH, device)
+
+    print("Generating predictions...")
+    # Get base model predictions
+    meta_features = ensemble.get_base_predictions(df)
+    prob_xgb = meta_features['prob_xgb'].values
+    prob_lstm = meta_features['prob_lstm'].values
+
+    # 1. Correlation Analysis
+    correlation = np.corrcoef(prob_xgb, prob_lstm)[0, 1]
+    print(f"\n[Correlation Analysis]")
+    print(f"Correlation between XGBoost and LSTM predictions: {correlation:.4f}")
+
+    # 2. Performance Comparison (Log Loss & Accuracy)
+    acc_xgb = accuracy_score(y, (prob_xgb > 0.5).astype(int))
+    ll_xgb = log_loss(y, prob_xgb)
+
+    acc_lstm = accuracy_score(y, (prob_lstm > 0.5).astype(int))
+    ll_lstm = log_loss(y, prob_lstm)
+
+    print(f"\n[Performance Comparison]")
+    print(f"XGBoost - Acc: {acc_xgb:.2%}, LogLoss: {ll_xgb:.4f}")
+    print(f"LSTM    - Acc: {acc_lstm:.2%}, LogLoss: {ll_lstm:.4f}")
+
+    # 3. Disagreement Analysis
+    # Where do they disagree?
+    pred_xgb = (prob_xgb > 0.5).astype(int)
+    pred_lstm = (prob_lstm > 0.5).astype(int)
+
+    disagreement_mask = pred_xgb != pred_lstm
+    disagreement_count = np.sum(disagreement_mask)
+    print(f"\n[Disagreement Analysis]")
+    print(f"Models disagree on {disagreement_count} / {len(df)} samples ({disagreement_count/len(df):.2%})")
+
+    if disagreement_count > 0:
+        # Who is right when they disagree?
+        disagreements = df[disagreement_mask].copy()
+        y_disagree = y[disagreement_mask]
+        pred_xgb_disagree = pred_xgb[disagreement_mask]
+        pred_lstm_disagree = pred_lstm[disagreement_mask]
+
+        xgb_correct = np.sum(pred_xgb_disagree == y_disagree)
+        lstm_correct = np.sum(pred_lstm_disagree == y_disagree)
+
+        print(f"In disagreement cases:")
+        print(f"  XGBoost correct: {xgb_correct} times")
+        print(f"  LSTM correct:    {lstm_correct} times")
+
+        # Show a few examples
+        print("\nExample Disagreements:")
+        disagreements['prob_xgb'] = prob_xgb[disagreement_mask]
+        disagreements['prob_lstm'] = prob_lstm[disagreement_mask]
+        disagreements['actual'] = y_disagree
+        print(disagreements[['round', 'tick', 'prob_xgb', 'prob_lstm', 'actual']].head(5))
+
+if __name__ == "__main__":
+    analyze_ensemble()
diff --git a/src/analysis/explain_model_shap.py b/src/analysis/explain_model_shap.py
new file mode 100644
index 0000000..e306404
--- /dev/null
+++ b/src/analysis/explain_model_shap.py
@@ -0,0 +1,95 @@
+import os
+import sys
+import pandas as pd
+import numpy as np
+import xgboost as xgb
+import shap
+import matplotlib.pyplot as plt
+
+# Add project root to path
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+
+from src.features.definitions import FEATURE_COLUMNS
+
+# Configuration
+MODEL_PATH = "models/clutch_model_v1.json"
+TEST_DATA_PATH = "data/processed/test_set.parquet"
+REPORT_DIR = "reports"
+
+def explain_model():
+    # 1. Ensure Report Directory Exists
+    if not os.path.exists(REPORT_DIR):
+        os.makedirs(REPORT_DIR)
+
+    # 2. Load Data
+    if not os.path.exists(TEST_DATA_PATH):
+        print(f"Error: Test data not found at {TEST_DATA_PATH}")
+        return
+
+    print(f"Loading test data from {TEST_DATA_PATH}...")
+    df = pd.read_parquet(TEST_DATA_PATH)
+    X = df[FEATURE_COLUMNS]
+    y = df['round_winner']
+
+    # 3. Load Model
+    if not os.path.exists(MODEL_PATH):
+        print(f"Error: Model not found at {MODEL_PATH}")
+        return
+
+    print(f"Loading XGBoost model from {MODEL_PATH}...")
+    model = xgb.XGBClassifier()
+    model.load_model(MODEL_PATH)
+
+    # 4. Calculate SHAP Values
+    print("Calculating SHAP values (this may take a moment)...")
+    explainer = shap.TreeExplainer(model)
+    shap_values = explainer.shap_values(X)
+
+    # 5. Generate Summary Plot
+    print(f"Generating SHAP Summary Plot...")
+    plt.figure(figsize=(10, 8))
+    shap.summary_plot(shap_values, X, show=False)
+    
+    summary_plot_path = os.path.join(REPORT_DIR, "shap_summary_v1.png")
+    plt.savefig(summary_plot_path, bbox_inches='tight', dpi=300)
+    plt.close()
+    print(f"Saved summary plot to: {summary_plot_path}")
+
+    # 6. Generate Bar Plot (Global Feature Importance)
+    print(f"Generating SHAP Bar Plot...")
+    plt.figure(figsize=(10, 8))
+    shap.summary_plot(shap_values, X, plot_type="bar", show=False)
+    
+    bar_plot_path = os.path.join(REPORT_DIR, "shap_importance_v1.png")
+    plt.savefig(bar_plot_path, bbox_inches='tight', dpi=300)
+    plt.close()
+    print(f"Saved importance plot to: {bar_plot_path}")
+
+    # 7. Interview Insights Generation
+    print("\n" + "="*50)
+    print("           DATA ANALYST INTERVIEW INSIGHTS           ")
+    print("="*50)
+    
+    # Calculate mean absolute SHAP values for importance
+    mean_abs_shap = np.mean(np.abs(shap_values), axis=0)
+    feature_importance = pd.DataFrame({
+        'feature': FEATURE_COLUMNS,
+        'importance': mean_abs_shap
+    }).sort_values('importance', ascending=False)
+
+    top_3 = feature_importance.head(3)
+    
+    print("When an interviewer asks: 'What drives your model's predictions?'")
+    print("You can answer based on this data:")
+    print(f"1. The most critical factor is '{top_3.iloc[0]['feature']}'.")
+    print(f"   (Impact Score: {top_3.iloc[0]['importance']:.4f})")
+    print(f"2. Followed by '{top_3.iloc[1]['feature']}' and '{top_3.iloc[2]['feature']}'.")
+    print("\nBusiness Interpretation:")
+    print("- If 'economy' features are top: The model confirms that money buys win rate.")
+    print("- If 'spatial' features are top: The model understands map control is key.")
+    print("- If 'status' (health/alive) features are top: The model relies on basic manpower advantage.")
+    print("-" * 50)
+    print(f"Check the visualizations in the '{REPORT_DIR}' folder to practice your storytelling.")
+
+if __name__ == "__main__":
+    explain_model()
\ No newline at end of file
diff --git a/src/inference/ensemble_framework.py b/src/inference/ensemble_framework.py
new file mode 100644
index 0000000..01e770c
--- /dev/null
+++ b/src/inference/ensemble_framework.py
@@ -0,0 +1,182 @@
+"""
+Ensemble Framework: XGBoost + LSTM Fusion
+=========================================
+This script demonstrates the framework for combining the static state analysis of XGBoost
+with the temporal trend analysis of LSTM to produce a robust final prediction.
+
+Methodology:
+1. Load both trained models (XGBoost .json, LSTM .pth).
+2. Prepare input data:
+   - XGBoost: Takes single frame features (24 dims).
+   - LSTM: Takes sequence of last 10 frames (10x24 dims).
+3. Generate independent probabilities: P_xgb, P_lstm.
+4. Fuse predictions using Weighted Averaging:
+   P_final = alpha * P_xgb + (1 - alpha) * P_lstm
+5. Evaluate performance on the test set.
+
+Usage:
+python src/inference/ensemble_framework.py
+"""
+
+import os
+import sys
+import numpy as np
+import pandas as pd
+import torch
+import xgboost as xgb
+from sklearn.metrics import accuracy_score, classification_report, log_loss
+
+# Ensure imports work
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+from src.training.models import ClutchLSTM
+from src.training.sequence_prep import create_sequences
+from src.features.definitions import FEATURE_COLUMNS
+
+# Configuration
+XGB_MODEL_PATH = "models/clutch_v1.model.json"
+LSTM_MODEL_PATH = "models/clutch_lstm_v1.pth"
+TEST_DATA_PATH = "data/processed/test_set.parquet" # The test set saved by train.py
+
+# Fusion Hyperparameter
+# 0.6 means we trust XGBoost slightly more (currently it has higher accuracy ~84% vs LSTM ~77%)
+ALPHA = 0.6 
+
+class ClutchEnsemble:
+    def __init__(self, xgb_path, lstm_path, device='cpu'):
+        self.device = device
+        
+        # Load XGBoost
+        print(f"Loading XGBoost from {xgb_path}...")
+        self.xgb_model = xgb.XGBClassifier()
+        self.xgb_model.load_model(xgb_path)
+        
+        # Load LSTM
+        print(f"Loading LSTM from {lstm_path}...")
+        # Need to know input_dim (24 features)
+        self.lstm_model = ClutchLSTM(input_dim=len(FEATURE_COLUMNS)).to(device)
+        self.lstm_model.load_state_dict(torch.load(lstm_path, map_location=device))
+        self.lstm_model.eval()
+        
+    def predict(self, df):
+        """
+        End-to-end prediction on a dataframe.
+        Note: This handles the complexity of alignment.
+        LSTM needs 10 frames, so the first 9 frames of each match cannot have LSTM predictions.
+        Strategy: Fallback to XGBoost for the first 9 frames.
+        """
+        # 1. XGBoost Predictions (Fast, parallel)
+        print("Generating XGBoost predictions...")
+        X_xgb = df[FEATURE_COLUMNS]
+        # predict_proba returns [prob_0, prob_1], we want prob_1 (CT Win?) 
+        # Wait, check mapping. train.py: T=0, CT=1. So index 1 is CT win probability.
+        # But wait, XGBoost might have different class order if not careful.
+        # Usually classes_ is [0, 1].
+        probs_xgb = self.xgb_model.predict_proba(X_xgb)[:, 1]
+        
+        # 2. LSTM Predictions (Sequential)
+        print("Generating LSTM predictions...")
+        # We need to create sequences. 
+        # Ideally we reuse sequence_prep logic but we need to keep the index aligned with df.
+        
+        # Initialize with NaN or fallback
+        probs_lstm = np.full(len(df), np.nan)
+        
+        # Group by match to avoid cross-match leakage in sequence creation
+        # We need to iterate and fill `probs_lstm`
+        
+        # For efficiency, let's extract sequences using the helper, but we need to know WHICH rows they correspond to.
+        # create_sequences in sequence_prep.py returns arrays, stripping index.
+        # Let's write a custom generator here that preserves alignment.
+        
+        seq_len = 10
+        inputs_list = []
+        indices_list = []
+        
+        grouped = df.groupby(['match_id', 'round'])
+        
+        for (match_id, round_num), group in grouped:
+            group = group.sort_values('tick')
+            data = group[FEATURE_COLUMNS].values
+            
+            if len(data) < seq_len:
+                continue
+                
+            for i in range(len(data) - seq_len + 1):
+                # Sequence ends at index i + seq_len - 1
+                # This index corresponds to the row we are predicting for
+                row_idx = group.index[i + seq_len - 1]
+                seq = data[i : i + seq_len]
+                
+                inputs_list.append(seq)
+                indices_list.append(row_idx)
+                
+        if len(inputs_list) > 0:
+            inputs_tensor = torch.FloatTensor(np.array(inputs_list)).to(self.device)
+            
+            with torch.no_grad():
+                # Batch prediction
+                # Depending on RAM, might need mini-batches. For <10k rows, full batch is fine.
+                outputs = self.lstm_model(inputs_tensor)
+                # outputs is [batch, 1] (Sigmoid)
+                lstm_preds = outputs.cpu().numpy().flatten()
+                
+            # Fill the array
+            probs_lstm[indices_list] = lstm_preds
+            
+        # 3. Fusion
+        print("Fusing predictions...")
+        final_probs = []
+        
+        for p_x, p_l in zip(probs_xgb, probs_lstm):
+            if np.isnan(p_l):
+                # Fallback to XGBoost if insufficient history (start of round)
+                final_probs.append(p_x)
+            else:
+                # Weighted Average
+                p_final = ALPHA * p_x + (1 - ALPHA) * p_l
+                final_probs.append(p_final)
+                
+        return np.array(final_probs)
+
+def main():
+    if not os.path.exists(TEST_DATA_PATH):
+        print(f"Test set not found at {TEST_DATA_PATH}. Please run training first.")
+        return
+        
+    print(f"Loading test set from {TEST_DATA_PATH}...")
+    df_test = pd.read_parquet(TEST_DATA_PATH)
+    
+    # Ground Truth
+    y_true = df_test['round_winner'].map({'T': 0, 'CT': 1}).values
+    
+    # Initialize Ensemble
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    ensemble = ClutchEnsemble(XGB_MODEL_PATH, LSTM_MODEL_PATH, device)
+    
+    # Predict
+    y_prob = ensemble.predict(df_test)
+    y_pred = (y_prob > 0.5).astype(int)
+    
+    # Evaluate
+    acc = accuracy_score(y_true, y_pred)
+    ll = log_loss(y_true, y_prob)
+    
+    print("\n" + "="*50)
+    print("           ENSEMBLE MODEL RESULTS           ")
+    print("="*50)
+    print(f"🔥 Final Accuracy: {acc:.2%}")
+    print(f"📉 Log Loss:       {ll:.4f}")
+    print("-" * 50)
+    print("Detailed Report:")
+    print(classification_report(y_true, y_pred, target_names=['T', 'CT']))
+    print("="*50)
+    
+    # Compare with standalone XGBoost for reference
+    # (Since we have the loaded model, let's just check quickly)
+    print("\n[Reference] Standalone XGBoost Performance:")
+    y_prob_xgb = ensemble.xgb_model.predict_proba(df_test[FEATURE_COLUMNS])[:, 1]
+    y_pred_xgb = (y_prob_xgb > 0.5).astype(int)
+    print(f"XGB Accuracy: {accuracy_score(y_true, y_pred_xgb):.2%}")
+
+if __name__ == "__main__":
+    main()
diff --git a/src/inference/stacking_ensemble.py b/src/inference/stacking_ensemble.py
new file mode 100644
index 0000000..dbfe27e
--- /dev/null
+++ b/src/inference/stacking_ensemble.py
@@ -0,0 +1,217 @@
+"""
+Stacking Ensemble Framework (Advanced Fusion)
+=============================================
+Beyond simple weighted averaging, this script implements 'Stacking' (Stacked Generalization).
+It trains a Meta-Learner (Logistic Regression) to intelligently combine the predictions 
+of the base models (XGBoost + LSTM) based on the current context (e.g., game time).
+
+Architecture:
+1. Base Layer: XGBoost, LSTM
+2. Meta Layer: Logistic Regression
+   Input: [Prob_XGB, Prob_LSTM, Game_Time, Team_Alive_Diff]
+   Output: Final Probability
+
+Why this is better:
+- It learns WHEN to trust which model (e.g., trust LSTM more in late-game).
+- It can correct systematic biases of base models.
+
+Usage:
+python src/inference/stacking_ensemble.py
+"""
+
+import os
+import sys
+import numpy as np
+import pandas as pd
+import torch
+import xgboost as xgb
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import accuracy_score, classification_report, log_loss
+from sklearn.model_selection import train_test_split
+
+# Ensure imports work
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+from src.training.models import ClutchAttentionLSTM
+from src.features.definitions import FEATURE_COLUMNS
+
+# Configuration
+XGB_MODEL_PATH = "models/clutch_model_v1.json"
+LSTM_MODEL_PATH = "models/clutch_attention_lstm_v1.pth"
+TEST_DATA_PATH = "data/processed/test_set.parquet"
+
+class StackingEnsemble:
+    def __init__(self, xgb_path, lstm_path, device='cpu'):
+        self.device = device
+        self.meta_learner = LogisticRegression()
+        self.is_fitted = False
+        
+        # Load Base Models
+        print(f"Loading Base Model: XGBoost...")
+        self.xgb_model = xgb.XGBClassifier()
+        self.xgb_model.load_model(xgb_path)
+        
+        print(f"Loading Base Model: LSTM...")
+        self.lstm_model = ClutchAttentionLSTM(input_dim=len(FEATURE_COLUMNS), hidden_dim=64, num_layers=2, dropout=0.5).to(device)
+        self.lstm_model.load_state_dict(torch.load(lstm_path, map_location=device))
+        self.lstm_model.eval()
+
+    def get_base_predictions(self, df):
+        """Generates features for the meta-learner."""
+        # Reset index to ensure positional indexing works for probs_lstm
+        df = df.reset_index(drop=True)
+        
+        # 1. XGBoost Probabilities
+        X_xgb = df[FEATURE_COLUMNS]
+        probs_xgb = self.xgb_model.predict_proba(X_xgb)[:, 1]
+        
+        # 2. LSTM Probabilities
+        probs_lstm = np.full(len(df), np.nan)
+        seq_len = 10
+        inputs_list = []
+        indices_list = []
+        
+        grouped = df.groupby(['match_id', 'round'])
+        
+        for (match_id, round_num), group in grouped:
+            group = group.sort_values('tick')
+            data = group[FEATURE_COLUMNS].values
+            
+            if len(data) < seq_len:
+                continue
+                
+            for i in range(len(data) - seq_len + 1):
+                row_idx = group.index[i + seq_len - 1]
+                seq = data[i : i + seq_len]
+                inputs_list.append(seq)
+                indices_list.append(row_idx)
+                
+        if len(inputs_list) > 0:
+            inputs_tensor = torch.FloatTensor(np.array(inputs_list)).to(self.device)
+            with torch.no_grad():
+                outputs = self.lstm_model(inputs_tensor)
+                lstm_preds = outputs.cpu().numpy().flatten()
+            probs_lstm[indices_list] = lstm_preds
+            
+        # Handle NaNs (start of rounds) - Fill with XGBoost prediction (trust base state)
+        # This is a crucial imputation step for the meta-learner
+        mask_nan = np.isnan(probs_lstm)
+        probs_lstm[mask_nan] = probs_xgb[mask_nan]
+        
+        # 3. Construct Meta-Features
+        # We add 'game_time' to let the meta-learner learn temporal weighting
+        # We add 'alive_diff' to let it know the complexity of the situation
+        meta_features = pd.DataFrame({
+            'prob_xgb': probs_xgb,
+            'prob_lstm': probs_lstm,
+            'game_time': df['game_time'].values,
+            'alive_diff': df['alive_diff'].values
+        })
+        
+        return meta_features
+
+    def fit_meta_learner(self, df, y):
+        """Train the Meta-Learner on a validation set."""
+        print("Generating meta-features for training...")
+        X_meta = self.get_base_predictions(df)
+        
+        print("Training Meta-Learner (Logistic Regression)...")
+        self.meta_learner.fit(X_meta, y)
+        self.is_fitted = True
+        
+        # Analyze learned weights
+        coefs = self.meta_learner.coef_[0]
+        print("\n[Meta-Learner Insights]")
+        print("How much does it trust each signal?")
+        print(f"  Weight on XGBoost:   {coefs[0]:.4f}")
+        print(f"  Weight on LSTM:      {coefs[1]:.4f}")
+        print(f"  Weight on GameTime:  {coefs[2]:.4f}")
+        print(f"  Weight on AliveDiff: {coefs[3]:.4f}")
+        print("(Positive weight = positive correlation with CT winning)\n")
+
+    def predict(self, df):
+        if not self.is_fitted:
+            raise ValueError("Meta-learner not fitted! Call fit_meta_learner first.")
+            
+        X_meta = self.get_base_predictions(df)
+        return self.meta_learner.predict_proba(X_meta)[:, 1]
+
+def main():
+    if not os.path.exists(TEST_DATA_PATH):
+        print("Test data not found.")
+        return
+        
+    print(f"Loading data from {TEST_DATA_PATH}...")
+    df = pd.read_parquet(TEST_DATA_PATH)
+    
+    # Target Mapping
+    # Data is already 0/1, so no need to map 'T'/'CT'
+    y = df['round_winner'].values
+    
+    # Split Data for Meta-Learning
+    # We need a 'Meta-Train' set to train the unifier, and 'Meta-Test' to evaluate it.
+    # Since we only have 2 matches in test_set, let's split by match_id if possible.
+    unique_matches = df['match_id'].unique()
+    
+    if len(unique_matches) >= 2:
+        # Split 50/50 by match
+        mid = len(unique_matches) // 2
+        meta_train_matches = unique_matches[:mid]
+        meta_test_matches = unique_matches[mid:]
+        
+        train_mask = df['match_id'].isin(meta_train_matches)
+        test_mask = df['match_id'].isin(meta_test_matches)
+        
+        df_meta_train = df[train_mask]
+        y_meta_train = y[train_mask]
+        
+        df_meta_test = df[test_mask]
+        y_meta_test = y[test_mask]
+        
+        print(f"Split: Meta-Train ({len(df_meta_train)} rows) | Meta-Test ({len(df_meta_test)} rows)")
+    else:
+        print("Not enough matches for match-level split. Using random split (Caution: Leakage).")
+        df_meta_train, df_meta_test, y_meta_train, y_meta_test = train_test_split(df, y, test_size=0.5, random_state=42)
+
+    # Initialize Ensemble
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    stacking_model = StackingEnsemble(XGB_MODEL_PATH, LSTM_MODEL_PATH, device)
+    
+    # 1. Train Meta-Learner
+    stacking_model.fit_meta_learner(df_meta_train, y_meta_train)
+    
+    # 2. Evaluate on Held-out Meta-Test set
+    print("Evaluating Stacking Ensemble...")
+    y_prob = stacking_model.predict(df_meta_test)
+    y_pred = (y_prob > 0.5).astype(int)
+    
+    acc = accuracy_score(y_meta_test, y_pred)
+    ll = log_loss(y_meta_test, y_prob)
+    
+    print("\n" + "="*50)
+    print("           STACKING ENSEMBLE RESULTS           ")
+    print("="*50)
+    print(f"Final Accuracy: {acc:.2%}")
+    print(f"Log Loss:       {ll:.4f}")
+    print("-" * 50)
+    print(classification_report(y_meta_test, y_pred, target_names=['T', 'CT']))
+    
+    # Baseline Comparison
+    print("="*50)
+    print("[Baselines on Meta-Test Set]")
+    
+    # XGBoost Baseline
+    X_test_xgb = df_meta_test[FEATURE_COLUMNS]
+    y_pred_xgb = stacking_model.xgb_model.predict(X_test_xgb)
+    acc_xgb = accuracy_score(y_meta_test, y_pred_xgb)
+    print(f"XGBoost Only: {acc_xgb:.2%}")
+
+    # LSTM Baseline
+    meta_features_test = stacking_model.get_base_predictions(df_meta_test)
+    probs_lstm = meta_features_test['prob_lstm'].values
+    y_pred_lstm = (probs_lstm > 0.5).astype(int)
+    acc_lstm = accuracy_score(y_meta_test, y_pred_lstm)
+    print(f"LSTM Only:    {acc_lstm:.2%}")
+    print("="*50)
+
+if __name__ == "__main__":
+    main()
diff --git a/src/training/models.py b/src/training/models.py
new file mode 100644
index 0000000..3095154
--- /dev/null
+++ b/src/training/models.py
@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Attention(nn.Module):
+    def __init__(self, hidden_dim):
+        super(Attention, self).__init__()
+        self.hidden_dim = hidden_dim
+        self.attn = nn.Linear(hidden_dim, 1)
+
+    def forward(self, x):
+        # x shape: (batch, seq, hidden)
+        
+        # Calculate attention scores
+        # scores shape: (batch, seq, 1)
+        scores = self.attn(x)
+        
+        # Softmax over sequence dimension
+        # weights shape: (batch, seq, 1)
+        weights = F.softmax(scores, dim=1)
+        
+        # Weighted sum
+        # context shape: (batch, hidden)
+        # element-wise multiplication broadcasted, then sum over seq
+        context = torch.sum(x * weights, dim=1)
+        
+        return context, weights
+
+class ClutchLSTM(nn.Module):
+    def __init__(self, input_dim, hidden_dim=64, num_layers=2, output_dim=1, dropout=0.2):
+        super(ClutchLSTM, self).__init__()
+        
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        
+        # LSTM Layer
+        # batch_first=True means input shape is (batch, seq, feature)
+        self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, 
+                            batch_first=True, dropout=dropout)
+        
+        # Fully Connected Layer
+        self.fc = nn.Linear(hidden_dim, output_dim)
+        
+        # Sigmoid activation for binary classification
+        self.sigmoid = nn.Sigmoid()
+        
+    def forward(self, x):
+        # x shape: (batch, seq, feature)
+        
+        # Initialize hidden state with zeros
+        # Using x.device ensures tensors are on the same device (CPU/GPU)
+        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(x.device)
+        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(x.device)
+        
+        # Forward propagate LSTM
+        # out shape: (batch, seq, hidden_dim)
+        out, _ = self.lstm(x, (h0, c0))
+        
+        # Decode the hidden state of the last time step
+        # out[:, -1, :] takes the last output of the sequence
+        out = self.fc(out[:, -1, :])
+        out = self.sigmoid(out)
+        return out
+
+class ClutchAttentionLSTM(nn.Module):
+    def __init__(self, input_dim, hidden_dim=128, num_layers=2, output_dim=1, dropout=0.3):
+        super(ClutchAttentionLSTM, self).__init__()
+        
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        
+        # 1. Input Layer Norm (Stabilizes training)
+        self.layer_norm = nn.LayerNorm(input_dim)
+        
+        # 2. LSTM (Increased hidden_dim for capacity)
+        self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, 
+                            batch_first=True, dropout=dropout, bidirectional=False)
+        
+        # 3. Attention Mechanism
+        self.attention = Attention(hidden_dim)
+        
+        # 4. Fully Connected Layers
+        self.fc1 = nn.Linear(hidden_dim, 64)
+        self.relu = nn.ReLU()
+        self.dropout = nn.Dropout(dropout)
+        self.fc2 = nn.Linear(64, output_dim)
+        
+        self.sigmoid = nn.Sigmoid()
+        
+    def forward(self, x):
+        # x: (batch, seq, feature)
+        x = self.layer_norm(x)
+        
+        # LSTM
+        # out: (batch, seq, hidden)
+        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(x.device)
+        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(x.device)
+        lstm_out, _ = self.lstm(x, (h0, c0))
+        
+        # Attention
+        # context: (batch, hidden)
+        context, attn_weights = self.attention(lstm_out)
+        
+        # MLP Head
+        out = self.fc1(context)
+        out = self.relu(out)
+        out = self.dropout(out)
+        out = self.fc2(out)
+        out = self.sigmoid(out)
+        
+        return out
diff --git a/src/training/sequence_prep.py b/src/training/sequence_prep.py
new file mode 100644
index 0000000..c563c0a
--- /dev/null
+++ b/src/training/sequence_prep.py
@@ -0,0 +1,114 @@
+"""
+Sequence Data Preparation for LSTM/GRU Models
+
+This script transforms the L2 frame-level data into L3 sequence data suitable for RNNs.
+Output: (Batch_Size, Sequence_Length, Num_Features)
+"""
+
+import os
+import sys
+import numpy as np
+import pandas as pd
+import logging
+import torch
+from torch.utils.data import Dataset, DataLoader
+
+# Ensure we can import from src
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+
+from src.training.train import load_data, preprocess_features
+from src.features.definitions import FEATURE_COLUMNS
+
+# Configuration
+SEQ_LEN = 10  # 10 frames * 2s/frame = 20 seconds of context
+DATA_DIR = "data/processed"
+OUTPUT_DIR = "data/sequences"
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+class ClutchSequenceDataset(Dataset):
+    def __init__(self, sequences, targets):
+        self.sequences = torch.FloatTensor(sequences)
+        self.targets = torch.FloatTensor(targets)
+        
+    def __len__(self):
+        return len(self.sequences)
+    
+    def __getitem__(self, idx):
+        return self.sequences[idx], self.targets[idx]
+
+def create_sequences(df, seq_len=10):
+    """
+    Creates sliding window sequences from the dataframe.
+    Assumes df is already sorted by match, round, and time.
+    """
+    sequences = []
+    targets = []
+    
+    # Group by Match and Round to ensure we don't sequence across boundaries
+    # Using 'tick' to sort, assuming it increases with time
+    grouped = df.groupby(['match_id', 'round'])
+    
+    match_ids = []  # To store match_id for each sequence
+
+    for (match_id, round_num), group in grouped:
+        group = group.sort_values('tick')
+        
+        data = group[FEATURE_COLUMNS].values
+        labels = group['round_winner'].values # 0 for T, 1 for CT (need to verify mapping)
+        
+        # Check if we have enough data for at least one sequence
+        if len(data) < seq_len:
+            continue
+            
+        # Create sliding windows
+        for i in range(len(data) - seq_len + 1):
+            seq = data[i : i + seq_len]
+            label = labels[i + seq_len - 1] # Label of the last frame in sequence
+            
+            sequences.append(seq)
+            targets.append(label)
+            match_ids.append(match_id)
+            
+    return np.array(sequences), np.array(targets), np.array(match_ids)
+
+def prepare_sequence_data(save=True):
+    if not os.path.exists(OUTPUT_DIR):
+        os.makedirs(OUTPUT_DIR)
+        
+    logging.info("Loading raw data...")
+    raw_df = load_data(DATA_DIR)
+    
+    logging.info("Preprocessing to frame-level features...")
+    df = preprocess_features(raw_df)
+    
+    logging.info(f"Frame-level df shape: {df.shape}")
+    logging.info(f"Columns: {df.columns.tolist()}")
+    logging.info(f"Round Winner unique values: {df['round_winner'].unique()}")
+    
+    # Ensure target is numeric
+    # Check if mapping is needed
+    if df['round_winner'].dtype == 'object':
+        logging.info("Mapping targets from T/CT to 0/1...")
+        df['round_winner'] = df['round_winner'].map({'T': 0, 'CT': 1})
+    
+    logging.info(f"Round Winner unique values after mapping: {df['round_winner'].unique()}")
+    df = df.dropna(subset=['round_winner'])
+    logging.info(f"df shape after dropna: {df.shape}")
+
+    logging.info(f"Creating sequences (Length={SEQ_LEN})...")
+    X, y, matches = create_sequences(df, SEQ_LEN)
+    
+    logging.info(f"Generated {len(X)} sequences.")
+    logging.info(f"Shape: {X.shape}")
+    
+    if save:
+        logging.info(f"Saving to {OUTPUT_DIR}...")
+        np.save(os.path.join(OUTPUT_DIR, "X_seq.npy"), X)
+        np.save(os.path.join(OUTPUT_DIR, "y_seq.npy"), y)
+        np.save(os.path.join(OUTPUT_DIR, "matches_seq.npy"), matches)
+        
+    return X, y, matches
+
+if __name__ == "__main__":
+    prepare_sequence_data()
diff --git a/src/training/train.py b/src/training/train.py
index 3a9f758..9d45e15 100644
--- a/src/training/train.py
+++ b/src/training/train.py
@@ -93,6 +93,9 @@ def preprocess_features(df):
     df['is_t'] = (df['team_num'] == 2).astype(int)
     df['is_ct'] = (df['team_num'] == 3).astype(int)
     
+    # Fill NA in 'is_alive' before conversion
+    df['is_alive'] = df['is_alive'].fillna(0)
+
     # Calculate player specific metrics
     df['t_alive'] = df['is_t'] * df['is_alive'].astype(int)
     df['ct_alive'] = df['is_ct'] * df['is_alive'].astype(int)
diff --git a/src/training/train_lstm.py b/src/training/train_lstm.py
new file mode 100644
index 0000000..406592d
--- /dev/null
+++ b/src/training/train_lstm.py
@@ -0,0 +1,207 @@
+"""
+Train Attention-LSTM Model for Clutch-IQ
+"""
+
+import os
+import sys
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.utils.data import DataLoader, TensorDataset
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import accuracy_score, log_loss, classification_report
+
+# Ensure we can import from src
+sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
+from src.training.models import ClutchAttentionLSTM
+
+# Config
+SEQ_DIR = os.path.join("data", "sequences")
+MODEL_DIR = "models"
+MODEL_PATH = os.path.join(MODEL_DIR, "clutch_attention_lstm_v1.pth")
+BATCH_SIZE = 32
+EPOCHS = 50
+LR = 0.001
+PATIENCE = 10
+
+class EarlyStopping:
+    def __init__(self, patience=5, min_delta=0):
+        self.patience = patience
+        self.min_delta = min_delta
+        self.counter = 0
+        self.best_loss = None
+        self.early_stop = False
+
+    def __call__(self, val_loss):
+        if self.best_loss is None:
+            self.best_loss = val_loss
+        elif val_loss > self.best_loss - self.min_delta:
+            self.counter += 1
+            if self.counter >= self.patience:
+                self.early_stop = True
+        else:
+            self.best_loss = val_loss
+            self.counter = 0
+
+def train_lstm():
+    if not os.path.exists(MODEL_DIR):
+        os.makedirs(MODEL_DIR)
+
+    # 1. Load Data
+    x_path = os.path.join(SEQ_DIR, "X_seq.npy")
+    y_path = os.path.join(SEQ_DIR, "y_seq.npy")
+    m_path = os.path.join(SEQ_DIR, "matches_seq.npy")
+    
+    if not os.path.exists(x_path) or not os.path.exists(y_path):
+        print(f"Data not found at {SEQ_DIR}. Please run src/training/sequence_prep.py first.")
+        return
+
+    print("Loading sequence data...")
+    X = np.load(x_path)
+    y = np.load(y_path)
+    
+    # Load match IDs if available, else warn and use random split
+    if os.path.exists(m_path):
+        matches = np.load(m_path)
+        print(f"Loaded match IDs. Shape: {matches.shape}")
+    else:
+        print("Warning: matches_seq.npy not found. Using random split (risk of leakage).")
+        matches = None
+    
+    print(f"Data Shape: X={X.shape}, y={y.shape}")
+    
+    # 2. Split
+    if matches is not None:
+        # GroupSplit
+        unique_matches = np.unique(matches)
+        print(f"Total unique matches: {len(unique_matches)}")
+        
+        # Shuffle matches
+        np.random.seed(42)
+        np.random.shuffle(unique_matches)
+        
+        n_train = int(len(unique_matches) * 0.8)
+        train_match_ids = unique_matches[:n_train]
+        test_match_ids = unique_matches[n_train:]
+        
+        print(f"Train matches: {len(train_match_ids)}, Test matches: {len(test_match_ids)}")
+        
+        train_mask = np.isin(matches, train_match_ids)
+        test_mask = np.isin(matches, test_match_ids)
+        
+        X_train, X_test = X[train_mask], X[test_mask]
+        y_train, y_test = y[train_mask], y[test_mask]
+    else:
+        # Stratify is important for imbalanced datasets
+        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
+    
+    print(f"Train set: {X_train.shape}, Test set: {X_test.shape}")
+    
+    # 3. Convert to PyTorch Tensors
+    train_data = TensorDataset(torch.from_numpy(X_train).float(), torch.from_numpy(y_train).float())
+    test_data = TensorDataset(torch.from_numpy(X_test).float(), torch.from_numpy(y_test).float())
+    
+    train_loader = DataLoader(train_data, batch_size=BATCH_SIZE, shuffle=True)
+    test_loader = DataLoader(test_data, batch_size=BATCH_SIZE)
+    
+    # 4. Model Setup
+    input_dim = X.shape[2] # Number of features
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"Training on {device}...")
+    
+    # Initialize Attention LSTM with higher dropout and lower complexity
+    model = ClutchAttentionLSTM(input_dim=input_dim, hidden_dim=64, num_layers=2, dropout=0.5).to(device)
+    
+    criterion = nn.BCELoss()
+    # Add weight decay for L2 regularization
+    optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=1e-4)
+    scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=3)
+    early_stopping = EarlyStopping(patience=PATIENCE, min_delta=0.0001)
+    
+    # 5. Train Loop
+    best_loss = float('inf')
+    
+    print("-" * 50)
+    print(f"{'Epoch':<6} | {'Train Loss':<12} | {'Val Loss':<12} | {'Val Acc':<10} | {'LR':<10}")
+    print("-" * 50)
+    
+    for epoch in range(EPOCHS):
+        model.train()
+        train_loss = 0.0
+        for inputs, labels in train_loader:
+            inputs, labels = inputs.to(device), labels.to(device)
+            labels = labels.unsqueeze(1) # (batch, 1)
+            
+            optimizer.zero_grad()
+            outputs = model(inputs)
+            loss = criterion(outputs, labels)
+            loss.backward()
+            optimizer.step()
+            
+            train_loss += loss.item() * inputs.size(0)
+            
+        train_loss /= len(train_loader.dataset)
+        
+        # Validation
+        model.eval()
+        val_loss = 0.0
+        correct = 0
+        total = 0
+        with torch.no_grad():
+            for inputs, labels in test_loader:
+                inputs, labels = inputs.to(device), labels.to(device)
+                labels = labels.unsqueeze(1)
+                outputs = model(inputs)
+                loss = criterion(outputs, labels)
+                val_loss += loss.item() * inputs.size(0)
+                
+                predicted = (outputs > 0.5).float()
+                total += labels.size(0)
+                correct += (predicted == labels).sum().item()
+        
+        val_loss /= len(test_loader.dataset)
+        val_acc = correct / total
+        
+        current_lr = optimizer.param_groups[0]['lr']
+        print(f"{epoch+1:<6} | {train_loss:.4f}       | {val_loss:.4f}       | {val_acc:.2%}    | {current_lr:.1e}")
+        
+        # Scheduler & Checkpointing
+        scheduler.step(val_loss)
+        
+        if val_loss < best_loss:
+            best_loss = val_loss
+            torch.save(model.state_dict(), MODEL_PATH)
+            # print(f"  -> Model saved (Val Loss: {val_loss:.4f})")
+            
+        # Early Stopping
+        early_stopping(val_loss)
+        if early_stopping.early_stop:
+            print("Early stopping triggered!")
+            break
+
+    print("Training Complete.")
+    
+    # 6. Final Evaluation
+    print(f"Loading best model from {MODEL_PATH}...")
+    model.load_state_dict(torch.load(MODEL_PATH))
+    model.eval()
+    
+    all_preds = []
+    all_labels = []
+    
+    with torch.no_grad():
+        for inputs, labels in test_loader:
+            inputs, labels = inputs.to(device), labels.to(device)
+            outputs = model(inputs)
+            preds = (outputs.squeeze() > 0.5).float()
+            all_preds.extend(preds.cpu().numpy())
+            all_labels.extend(labels.cpu().numpy())
+            
+    print("-" * 50)
+    print("Detailed Classification Report (Best Model):")
+    print(classification_report(all_labels, all_preds, target_names=['T (Terrorist)', 'CT (Counter-Terrorist)']))
+    print("="*50)
+
+if __name__ == "__main__":
+    train_lstm()