diff --git a/scripts/data_platform/corpus_download.py b/scripts/data_platform/corpus_download.py
index 47437a4..42a4eb6 100644
--- a/scripts/data_platform/corpus_download.py
+++ b/scripts/data_platform/corpus_download.py
@@ -672,8 +672,17 @@ def _filter_new(domain, norm_rows, ledger, recent):
     keys = [_row_key(domain, r) for r in norm_rows]
     known_l = ledger.has_any(keys)
     known_r = recent.has_any(keys)
-    return [r for r, kl, kr in zip(norm_rows, known_l, known_r)
-            if not kl and not kr]
+    # 批内去重(09-07 首跑实锤巨潮单响应同 id 两条: 整行相同/时间差1秒):
+    # has_any 只对账已有键,批内重复须在此截留,否则双双落盘成重键行。
+    seen = set()
+    out = []
+    for r, kl, kr in zip(norm_rows, known_l, known_r):
+        k = _row_key(domain, r)
+        if kl or kr or k in seen:
+            continue
+        seen.add(k)
+        out.append(r)
+    return out
 
 
 def _absorb_new(domain, norm_rows, ledgers, recents):
diff --git a/tests/data_platform/test_corpus_download.py b/tests/data_platform/test_corpus_download.py
index b60bd3e..b3b1386 100644
--- a/tests/data_platform/test_corpus_download.py
+++ b/tests/data_platform/test_corpus_download.py
@@ -218,6 +218,29 @@ def test_commit_batches_data_before_markers(root, monkeypatch):
     assert cd.IdLedger("ann_meta").has_any(["m1"]) == [True]
 
 
+def test_absorb_new_dedups_within_batch(root):
+    """批内去重(09-07 首跑实锤 ann_meta 3 行重键, verify_corpus FAIL):
+    巨潮单次响应同 id 可出现两次——①整行完全相同 ②ann_time 差 1 秒(目录挂两条)
+    ——都必须只落一行。根因: _filter_new 对整批先算 known 再统一 add,批内重复全放行。"""
+    import pandas as pd
+    ledgers = {"ann_meta": cd.IdLedger("ann_meta")}
+    recents = {"ann_meta": cd.RecentIndex("ann_meta", days=7)}
+    cd._DAY_BUFFERS.clear()
+    norm = [cd.norm_ann_row(r, "2026-09-07") for r in (
+        _ann_raw(aid="1225400726"),                        # 完全相同 ×2
+        _ann_raw(aid="1225400726"),
+        _ann_raw(aid="1225009396", ts_ms=1773616809000),   # 时间戳差 1 秒
+        _ann_raw(aid="1225009396", ts_ms=1773616810000),
+        _ann_raw(aid="1225487003"),                        # 正常单条
+    )]
+    new = cd._absorb_new("ann_meta", norm, ledgers, recents)
+    assert len(new) == 3
+    cd.flush_domains()
+    today = dt.date.today().isoformat()
+    df = pd.read_parquet(root / "ann_meta" / f"dt={today}" / "part-0.parquet")
+    assert len(df) == 3 and df.announcement_id.is_unique
+
+
 def test_reconcile_ledgers_heals_crash_gap(root):
     """启动对账自愈(09-07 巡检实锤旧代码丢内存账本 ann 差4.5k/news 差9.9w):
     分区有而账本无 → 启动补齐;账本⊇近30日分区永久成立。"""