diff --git a/core/repair_jobs/missing_lyrics.py b/core/repair_jobs/missing_lyrics.py
index 4cf02a93..1cf9d95d 100644
--- a/core/repair_jobs/missing_lyrics.py
+++ b/core/repair_jobs/missing_lyrics.py
@@ -114,8 +114,13 @@ class MissingLyricsJob(RepairJob):
             # Option A: only flag tracks LRClib actually has lyrics for. An
             # instrumental returns nothing here and is silently skipped (never
             # re-flagged on future scans).
+            # tracks.duration is stored in MILLISECONDS (schema) — LRClib's
+            # exact-match wants SECONDS, so convert (else exact-match never
+            # hits and it silently falls back to a fuzzier search).
             try:
-                duration_s = int(duration) if duration else None
+                duration_s = int(int(duration) / 1000) if duration else None
+                if duration_s is not None and duration_s <= 0:
+                    duration_s = None
             except (TypeError, ValueError):
                 duration_s = None
             try:
diff --git a/tests/test_missing_lyrics_job.py b/tests/test_missing_lyrics_job.py
index fe7038e2..3adfbd9c 100644
--- a/tests/test_missing_lyrics_job.py
+++ b/tests/test_missing_lyrics_job.py
@@ -112,6 +112,23 @@ def test_scan_flags_only_tracks_with_available_lyrics(tmp_path, monkeypatch):
     assert findings[0]["details"]["track_title"] == "Song"   # the instrumental was skipped
 
 
+def test_scan_converts_duration_ms_to_seconds(tmp_path, monkeypatch):
+    # tracks.duration is milliseconds; LRClib wants seconds. The scan must
+    # convert before querying (215000ms → 215s) and store seconds in the finding.
+    t1 = tmp_path / "song.flac"; t1.write_bytes(b"x")
+    rows = [(1, "Song", "Artist", "Album", str(t1), 215000)]   # 215000 ms = 215 s
+    seen = {}
+    fake_client = SimpleNamespace(
+        api=object(),
+        has_remote_lyrics=lambda title, artist, album, dur: seen.update(dur=dur) or True)
+    monkeypatch.setattr("core.lyrics_client.lyrics_client", fake_client)
+
+    findings = []
+    MissingLyricsJob().scan(_ctx(_DB(rows), findings))
+    assert seen["dur"] == 215                       # converted to seconds
+    assert findings[0]["details"]["duration"] == 215
+
+
 def test_scan_skips_tracks_that_already_have_lrc(tmp_path, monkeypatch):
     t1 = tmp_path / "song.flac"; t1.write_bytes(b"x")
     (tmp_path / "song.lrc").write_text("[00:01]hi")   # already has lyrics