diff --git a/scripts/diag_generate.py b/scripts/diag_generate.py
new file mode 100644
index 0000000..cc96203
--- /dev/null
+++ b/scripts/diag_generate.py
@@ -0,0 +1,49 @@
+"""层 1 关账判据 3：训练后 checkpoint 能被 from_pretrained 加载并生成通顺解答。
+
+远程运行（CPU 即可，0.6B 生成 512 token 约 1-2 分钟）：
+    python -u scripts/diag_generate.py
+"""
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from ars_opd.configs import SFTConfig
+from ars_opd.data import load_sft_dataset
+
+MODEL_DIR = "/data/zym/outputs/sft_qwen3-0.6b_dapo1k"  # 正式 1 epoch 的产物
+
+cfg = SFTConfig(
+    dataset_path="data/dapo-math-17k-unique.parquet",
+    output_dir="/tmp/diag",
+    subset_size=1000,
+    seed=42,
+    # 不挂 teacher 解答：只取题目做推理输入
+)
+ds = load_sft_dataset(cfg)
+
+tok = AutoTokenizer.from_pretrained(MODEL_DIR)
+model = AutoModelForCausalLM.from_pretrained(MODEL_DIR, dtype=torch.float32)
+model.eval()
+
+# 取子集第 900+ 行附近的题（训练时见过，此处只验"会不会说话"不验泛化）
+for i in (900, 950):
+    prompt = tok.apply_chat_template(
+        ds[i]["messages"],
+        tokenize=False,
+        add_generation_prompt=True,
+        enable_thinking=False,  # 必须与训练取值一致（docs/02 §2.3 边界契约）
+    )
+    inputs = tok(prompt, return_tensors="pt", add_special_tokens=False)
+    with torch.no_grad():
+        out = model.generate(
+            **inputs, max_new_tokens=512, do_sample=False, temperature=None, top_p=None
+        )
+    completion = tok.decode(out[0][inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+    print(f"===== 样本 {i} 题目 =====")
+    print(ds[i]["messages"][-1]["content"][120:280], "…")
+    print(f"----- 生成（前 600 字符）-----")
+    print(completion[:600])
+    print()
+
+print("判读：应为步骤化数学解答（markdown 风格、以 Answer: 行收尾的倾向）；"
+      "乱码/复读/空输出 = 不通过。", flush=True)