diff --git a/core/matching/audio_verification.py b/core/matching/audio_verification.py
new file mode 100644
index 00000000..45ff48e6
--- /dev/null
+++ b/core/matching/audio_verification.py
@@ -0,0 +1,59 @@
+"""Shared audio-verification decision core (pure; no file/DB I/O).
+
+Single source of truth for normalization + the PASS/SKIP/FAIL decision used by
+BOTH import-time verification (``core/acoustid_verification.py``) and the library
+scan (``core/repair_jobs/acoustid_scanner.py``). Historically each path had its
+own ``_normalize`` and decision branches that drifted apart and produced
+inconsistent results (a correct cross-script anime-OST track passed at import but
+was false-flagged by the scan). Centralising the decision here means the
+thresholds, normalization, alias-aware comparison, cross-script handling, version
+gate and duration guard are defined exactly once.
+"""
+
+import re
+from difflib import SequenceMatcher
+
+# Thresholds — the single definition both paths share.
+MIN_ACOUSTID_SCORE = 0.80       # Minimum fingerprint score to trust a match.
+TITLE_MATCH_THRESHOLD = 0.70    # Title similarity to consider a match.
+ARTIST_MATCH_THRESHOLD = 0.60   # Artist similarity to consider a match.
+
+
+def normalize(text: str) -> str:
+    """Normalize a title/artist for comparison.
+
+    lowercase; strip ``()`` / ``[]`` / ``<>`` annotations (version tags,
+    performer credits like ``<Vocal: MIKA KOBAYASHI>``); strip trailing
+    version / featuring tags; KEEP CJK characters (``\\w`` is unicode-aware) so
+    Japanese/Chinese/Korean titles produce a comparable form instead of an empty
+    string; collapse whitespace.
+    """
+    if not text:
+        return ""
+    s = text.lower().strip()
+    # Annotations that are metadata, not core identity.
+    s = re.sub(r'\s*\([^)]*\)', '', s)
+    s = re.sub(r'\s*\[[^\]]*\]', '', s)
+    s = re.sub(r'\s*<[^>]*>', '', s)
+    # Trailing featuring / version tags.
+    s = re.sub(r'\s+(?:feat\.?|ft\.?|featuring)\s+.*$', '', s, flags=re.IGNORECASE)
+    s = re.sub(
+        r'\s*-\s*(?:vocal|instrumental|acoustic|live|remix|cover|clean|explicit|'
+        r'radio\s*edit|original\s*mix|extended\s*mix|club\s*mix)\s*$',
+        '', s, flags=re.IGNORECASE,
+    )
+    s = re.sub(r'\s*-\s*from\s+.+$', '', s, flags=re.IGNORECASE)
+    # Drop remaining punctuation but keep word chars (incl. CJK) + spaces.
+    s = re.sub(r'[^\w\s]', '', s)
+    s = re.sub(r'\s+', ' ', s).strip()
+    return s
+
+
+def similarity(a: str, b: str) -> float:
+    """Similarity (0.0–1.0) between two strings after normalization."""
+    na, nb = normalize(a), normalize(b)
+    if not na or not nb:
+        return 0.0
+    if na == nb:
+        return 1.0
+    return SequenceMatcher(None, na, nb).ratio()
diff --git a/tests/matching/test_audio_verification_core.py b/tests/matching/test_audio_verification_core.py
new file mode 100644
index 00000000..a9dd2e85
--- /dev/null
+++ b/tests/matching/test_audio_verification_core.py
@@ -0,0 +1,22 @@
+"""Shared audio-verification decision core: normalize() + evaluate().
+
+One place for normalization + the PASS/SKIP/FAIL decision used by BOTH import-time
+verification and the library AcoustID scan, so the two paths can't drift apart.
+"""
+
+from core.matching.audio_verification import normalize
+
+
+def test_normalize_strips_paren_bracket_angle_and_keeps_cjk():
+    assert normalize("澤野弘之 <Vocal: MIKA KOBAYASHI>") == "澤野弘之"
+    assert normalize("Clarity (Live at X) [Remastered]") == "clarity"
+    assert normalize("Attack on Titan <TV Size>") == "attack on titan"
+
+
+def test_normalize_strips_version_and_featuring():
+    assert normalize("In My Feelings - Instrumental") == "in my feelings"
+    assert normalize("Song feat. Someone") == "song"
+
+
+def test_normalize_keeps_plain_text():
+    assert normalize("Sawano Hiroyuki") == "sawano hiroyuki"