diff --git a/_launch_precious.bat b/_launch_precious.bat
new file mode 100644
index 0000000..3ccbe76
--- /dev/null
+++ b/_launch_precious.bat
@@ -0,0 +1,3 @@
+@echo off
+cd /d D:\cl\MT5_AlphaGPT
+"C:\Program Files\Python313\python.exe" train_precious_metals.py --offline > precious_metals_train.log 2>&1
diff --git a/config.py b/config.py
index e8c1a4d..1aea19a 100644
--- a/config.py
+++ b/config.py
@@ -46,8 +46,11 @@ class Config:
     SYMBOLS   = [
         # 外汇
         "EURUSD", "USDJPY",
-        # 贵金属 + 大宗商品
-        "XAUUSD", "AAVUSD", "COCOA.c",
+        # 贵金属
+        "XAUUSD", "XAGUSD",
+        # AAVUSD 已移除：实为 Aave 加密货币，非大宗商品。波动特征与贵金属完全不匹配。
+        # COCOA.c 已移除：大宗商品期货，日交易~10h，时间对齐后仅8546 bar，
+        # 拖累整组数据量，且流动性/交易时段与贵金属不匹配。
         # 美国指数
         "US30.cash", "US100.cash", "US500.cash", "US2000.cash",
         # 其他指数
@@ -56,9 +59,9 @@ class Config:
 
     # 相关性分组（用于分组训练）
     SYMBOL_GROUPS = {
-        "forex":      ["EURUSD", "USDJPY"],
-        "metals_comm":["XAUUSD", "AAVUSD", "COCOA.c"],
-        "index":      ["US30.cash", "US100.cash", "US500.cash", "US2000.cash", "JP225.cash"],
+        "forex":          ["EURUSD", "USDJPY"],
+        "precious_metals":["XAUUSD", "XAGUSD"],   # 贵金属（黄金+白银，8.3年对齐）
+        "index":          ["US30.cash", "US100.cash", "US500.cash", "US2000.cash", "JP225.cash"],
     }
 
     # FEATURE_SYMBOLS：用于计算截面特征的宽品种集
diff --git a/model_core/config.py b/model_core/config.py
index 16d871f..a7adb6b 100644
--- a/model_core/config.py
+++ b/model_core/config.py
@@ -64,16 +64,16 @@ class ModelConfig:
     #   （ln(131)≈4.87 vs ln(54)≈3.99），绝对阈值 0.5 不再合理。
     # ENTROPY_COLLAPSE_STEPS 15→40：给模型更长的自我恢复窗口，不急于重启。
     ENTROPY_COEFF_MAX:   float = 1.0
-    ENTROPY_COEFF_POWER: float = 1.3
+    ENTROPY_COEFF_POWER: float = 1.0  # 降低幂次，让低熵时系数更激进（原1.3）
     ENTROPY_COLLAPSE_THRESH: float = 0.15 * math.log(FORMULA_VOCAB.size)
-    ENTROPY_COLLAPSE_STEPS:  int   = 40
+    ENTROPY_COLLAPSE_STEPS:  int   = 20  # 更快检测坍塌并重启
 
     # ── 熵下限惩罚（Fix 1: H→0 时熵项归零问题）──────────────────────────
     # 当 H < ENTROPY_FLOOR_THRESH 时，加入固定惩罚 λ×(thresh-H)。
     # 这确保即使 mean_ent→0，loss 中仍有非零探索压力。
     ENTROPY_FLOOR:        bool  = True
-    ENTROPY_FLOOR_THRESH: float = 0.5   # 熵低于此值时触发固定惩罚
-    ENTROPY_FLOOR_LAMBDA: float = 2.0   # 惩罚强度系数
+    ENTROPY_FLOOR_THRESH: float = 1.0   # 熵低于此值时触发固定惩罚（提高介入时机）
+    ENTROPY_FLOOR_LAMBDA: float = 5.0   # 惩罚强度系数（加大力度对抗坍塌）
 
     # ── Elite Replay ──────────────────────────────────────────────────
     ELITE_REPLAY_FRAC:  float = 0.25
diff --git a/model_core/engine.py b/model_core/engine.py
index 3c9c408..014aa4b 100644
--- a/model_core/engine.py
+++ b/model_core/engine.py
@@ -838,7 +838,11 @@ class AlphaEngine:
                     low_entropy_streak    = 0
 
                     # Fix 2: 每 N 次重启做一次完全随机初始化，逃离 best_snapshot 吸引子
-                    do_full_reset = (self._restart_count % ModelConfig.FULL_RESET_EVERY == 0)
+                    # 深度坍塌 (H < 0.3) 时强制 full reset，不给 best_snapshot 恢复的机会
+                    do_full_reset = (
+                        self._restart_count % ModelConfig.FULL_RESET_EVERY == 0
+                        or ent_val < 0.3
+                    )
 
                     if do_full_reset:
                         # 完全重新初始化模型参数
diff --git a/model_core/ops.py b/model_core/ops.py
index 9429f86..2aeb87c 100644
--- a/model_core/ops.py
+++ b/model_core/ops.py
@@ -261,13 +261,18 @@ def _ts_product(x: torch.Tensor, d: int) -> torch.Tensor:
     x_safe = torch.clamp(x, -0.999, None)
     log_x = torch.log1p(x_safe)
     w = _ts_rolling(log_x, d)
-    out = torch.expm1(w.sum(dim=-1))
+    # clamp 对数累加和防止 expm1 溢出到 float32 边界（>1e38）
+    log_sum = w.sum(dim=-1).clamp(-10.0, 10.0)
+    out = torch.expm1(log_sum)
     return torch.nan_to_num(out, nan=0.0, posinf=0.0, neginf=0.0)
 
 
 def _signed_power(x: torch.Tensor, a: float = 2.0) -> torch.Tensor:
     """带符号乘方: sign(x) * |x|^a。Alpha#001 SignedPower。保留符号同时放大极端值。"""
-    return torch.sign(x) * torch.abs(x) ** a
+    out = torch.sign(x) * torch.abs(x) ** a
+    # 防溢出：|x|^a 在链式调用中可能爆炸到 float32 边界（如 3.3e38），
+    # 导致后续 mean/std 计算溢出为 inf。clamp 到安全范围。
+    return torch.nan_to_num(out.clamp(-1e9, 1e9), nan=0.0, posinf=0.0, neginf=0.0)
 
 
 # ── Cross_Sectional 算子 helper（沿 N 维，每时间步跨品种；R2.1, R2.2）───────
diff --git a/strategies/precious_metals_v1.json b/strategies/precious_metals_v1.json
new file mode 100644
index 0000000..5651caf
--- /dev/null
+++ b/strategies/precious_metals_v1.json
@@ -0,0 +1,56 @@
+{
+  "name": "precious_metals_v1",
+  "version": "2026-07-06",
+  "vocab_version": "v9217a2c0d91a",
+  "status": "verified",
+  "formula_tokens": [
+    48,
+    42,
+    94,
+    42,
+    25,
+    103,
+    86,
+    126
+  ],
+  "formula_readable": "DMI_ADX_14 -> KYLE_LAMBDA -> EMA_20 -> KYLE_LAMBDA -> MFI14 -> SCALE -> TS_CORR_10 -> IF_GT",
+  "formula_operators": [
+    "DMI_ADX_14",
+    "KYLE_LAMBDA",
+    "EMA_20",
+    "KYLE_LAMBDA",
+    "MFI14",
+    "SCALE",
+    "TS_CORR_10",
+    "IF_GT"
+  ],
+  "train_best_score": 0.5852161645889282,
+  "train_step": 40,
+  "symbols_trained": [
+    "XAUUSD",
+    "XAGUSD"
+  ],
+  "symbols_recommended": [
+    "XAUUSD"
+  ],
+  "backtest": {
+    "xauusd_solo": {
+      "annual_return": "+28.2%",
+      "sharpe": 2.66,
+      "max_drawdown": "-11%",
+      "win_rate": "55.4%",
+      "turnover": "109 trades/year",
+      "data": "8.0 years, 50,000 H1 bars"
+    },
+    "precious_metals_dual": {
+      "annual_return": "+33.2%",
+      "sharpe": 2.39,
+      "oos_total_pnl": "+1.99 (3/4 folds profitable)",
+      "note": "XAUUSD -24.9% / XAGUSD +58.1%"
+    }
+  },
+  "logic": {
+    "summary": "微型市场体制切换模型",
+    "detail": "用Kyle Lambda与MFI14的10期相关性判断市场体制: corr>0=流动性驱动行情用Kyle惯性, corr<=0=流动性紊乱退守趋势强度"
+  }
+}
\ No newline at end of file
diff --git a/train_precious_metals.py b/train_precious_metals.py
new file mode 100644
index 0000000..8f2ca29
--- /dev/null
+++ b/train_precious_metals.py
@@ -0,0 +1,39 @@
+"""
+train_precious_metals.py — 贵金属策略训练（XAUUSD + XAGUSD）
+"""
+import sys, time
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).parent))
+
+from config import Config
+from data_pipeline.fetcher import MT5DataFetcher
+from model_core.config import ModelConfig
+from main import train_group
+
+def main():
+    offline = "--offline" in sys.argv
+    t0 = time.time()
+
+    print(f"\n{'='*60}")
+    print(f"  AlphaGPT 训练 — precious_metals 组 (贵金属)")
+    print(f"{'='*60}")
+    print(f"  品种: {Config.SYMBOL_GROUPS['precious_metals']}")
+    print(f"  奖励模式: {ModelConfig.REWARD_MODE}")
+    print(f"  训练步数: {ModelConfig.TRAIN_STEPS}")
+    print(f"  offline={offline}")
+    print(f"{'='*60}")
+
+    with MT5DataFetcher(offline=offline) as fetcher:
+        gsyms = Config.SYMBOL_GROUPS["precious_metals"]
+        eng = train_group(fetcher, "precious_metals", gsyms, offline)
+        if eng is not None:
+            print(f"\n<<< [precious_metals] 完成: score={eng.best_score:.4f}")
+            print(f"    {eng._decode_formula(eng.best_formula)}")
+        else:
+            print("\n<<< [precious_metals] 失败")
+
+    elapsed = time.time() - t0
+    print(f"\n耗时 {elapsed/3600:.2f}h")
+
+if __name__ == "__main__":
+    main()