diff --git a/src/ocr_sprint/pipeline/extract/personnel_text.py b/src/ocr_sprint/pipeline/extract/personnel_text.py
index 4360a80..5e37984 100644
--- a/src/ocr_sprint/pipeline/extract/personnel_text.py
+++ b/src/ocr_sprint/pipeline/extract/personnel_text.py
@@ -50,8 +50,15 @@ _RE_RANK_NRP_LINE = re.compile(
 # line in tabular layouts.
 _RE_ROW_NUMBER = re.compile(r"^\s*(\d{1,3})\s*[.)]\s*$")
 # Lines that should never be interpreted as a personnel name. These are
-# section headers, OCR garbage anchors, and column header tokens.
-_NAME_BLOCKLIST_PREFIXES: tuple[str, ...] = (
+# section headers, OCR garbage anchors, and column header tokens. We match
+# them with a *word-boundary* regex (built from this list) rather than a
+# bare ``startswith`` check, because short tokens like ``"NO"`` and
+# ``"KET"`` would otherwise reject perfectly valid Indonesian names
+# (e.g. ``"NOVA SARI"``, ``"NOOR HIDAYAT"``, ``"KETUT WARDANA"`` — the
+# latter being an extremely common Balinese birth-order name).
+_NAME_BLOCKLIST_TOKENS: tuple[str, ...] = (
+    "PADA TANGGAL",  # multi-word entries first so they win the alternation
+    "SURAT PERINTAH",
     "DASAR",
     "PERIHAL",
     "PERTIMBANGAN",
@@ -60,7 +67,6 @@ _NAME_BLOCKLIST_PREFIXES: tuple[str, ...] = (
     "UNTUK",
     "TEMBUSAN",
     "DIKELUARKAN",
-    "PADA TANGGAL",
     "SELESAI",
     "DAFTAR",
     "LAMPIRAN",
@@ -71,7 +77,6 @@ _NAME_BLOCKLIST_PREFIXES: tuple[str, ...] = (
     "RESOR",
     "SEKTOR",
     "MABES",
-    "SURAT PERINTAH",
     "NRP",
     "NIP",
     "PANGKAT",
@@ -81,6 +86,10 @@ _NAME_BLOCKLIST_PREFIXES: tuple[str, ...] = (
     "KET",
     "NO",
 )
+_RE_NAME_BLOCKLIST = re.compile(
+    r"^(?:" + "|".join(re.escape(tok) for tok in _NAME_BLOCKLIST_TOKENS) + r")\b",
+    re.IGNORECASE,
+)
 # A name should look like a name: mostly letters, common punctuation, and
 # at least one alphabetic character. Pure-numeric or pure-symbol lines are
 # rejected.
@@ -92,10 +101,8 @@ def _is_plausible_name(line: str) -> bool:
     stripped = line.strip()
     if not stripped or not _RE_NAME_OK.search(stripped):
         return False
-    upper = stripped.upper()
-    for prefix in _NAME_BLOCKLIST_PREFIXES:
-        if upper.startswith(prefix):
-            return False
+    if _RE_NAME_BLOCKLIST.match(stripped):
+        return False
     if _RE_ROW_NUMBER.match(stripped):
         return False
     if _RE_RANK_NRP_LINE.search(stripped):
@@ -122,8 +129,7 @@ def _following_jabatan(lines: list[str], idx: int) -> str | None:
             break
         if _RE_ROW_NUMBER.match(candidate):
             break
-        upper = candidate.upper()
-        if any(upper.startswith(p) for p in _NAME_BLOCKLIST_PREFIXES):
+        if _RE_NAME_BLOCKLIST.match(candidate):
             break
         parts.append(candidate)
     if not parts:
diff --git a/tests/unit/test_personnel_text_fallback.py b/tests/unit/test_personnel_text_fallback.py
index 95cdd26..884f99c 100644
--- a/tests/unit/test_personnel_text_fallback.py
+++ b/tests/unit/test_personnel_text_fallback.py
@@ -92,6 +92,37 @@ class TestExtractPersonnelFromText:
         text = "Some line\nFAKERANK / 12345678\nanother line"
         assert extract_personnel_from_text(text) == []
 
+    def test_does_not_drop_indonesian_names_starting_with_no_or_ket(self) -> None:
+        # Regression: 'NO' / 'KET' are legitimate column header tokens but
+        # also prefix common Indonesian names (KETUT, NOVA, NOOR). The
+        # blocklist must use word boundaries, not a raw startswith check.
+        text = (
+            "DAFTAR PERSONIL\n"
+            "1.\n"
+            "KETUT WARDANA\n"
+            "AIPTU / 11111111\n"
+            "JABATAN A\n"
+            "2.\n"
+            "NOVA SARI\n"
+            "BRIPTU / 22222222\n"
+            "JABATAN B\n"
+            "3.\n"
+            "NOOR HIDAYAT\n"
+            "BRIPDA / 33333333\n"
+            "JABATAN C\n"
+        )
+        rows = extract_personnel_from_text(text)
+        names = [r.nama for r in rows]
+        assert names == ["KETUT WARDANA", "NOVA SARI", "NOOR HIDAYAT"]
+
+    def test_still_blocks_bare_column_header_tokens(self) -> None:
+        # Word-boundary fix must still reject the actual column-header
+        # rows that motivated the blocklist in the first place.
+        text = "NO\nNAMA\nPANGKAT / NRP\nJABATAN\nKET\n1.\nREAL NAME\nAIPTU / 12345678\n"
+        rows = extract_personnel_from_text(text)
+        assert len(rows) == 1
+        assert rows[0].nama == "REAL NAME"
+
 
 class TestIsLowQuality:
     def test_empty_list_is_low_quality(self) -> None: