diff --git a/pixi.toml b/pixi.toml
index e90e1ca..ba9d3d6 100644
--- a/pixi.toml
+++ b/pixi.toml
@@ -11,4 +11,4 @@ test = "mojo run -I src test/test_captions.mojo"
 demo = "mojo run -I src examples/clip_transcript.mojo test/data/sample.vtt 0 20000"
 
 [dependencies]
-mojo = ">=1.0.0b3,<2"
+mojo = ">=1.0.0b3.dev0,<2"
diff --git a/src/captions/captions.mojo b/src/captions/captions.mojo
index c05fdcc..55abb01 100644
--- a/src/captions/captions.mojo
+++ b/src/captions/captions.mojo
@@ -176,6 +176,24 @@ def _parse_timing(line: String, mut start_ms: Int, mut end_ms: Int) raises:
     end_ms = _parse_ts(right)
 
 
+def _is_timing_line(line: String) -> Bool:
+    """Whether `line` is a genuine `timestamp --> timestamp` timing line.
+
+    Cue text may legitimately contain `-->` (e.g. "map --> filter", or a
+    Unicode-arrow gloss), so a bare `-->` substring is not enough to mark
+    a line as a cue boundary; both sides must parse as timestamps.
+    """
+    if line.find("-->") == -1:
+        return False
+    var start_ms = 0
+    var end_ms = 0
+    try:
+        _parse_timing(line, start_ms, end_ms)
+    except:
+        return False
+    return True
+
+
 def _strip_voice_tags(text: String, mut speaker: String) -> String:
     """Remove `<v ...>` / `</v>` markup; record the first voice's name."""
     if text.find("<v") == -1:
@@ -285,7 +303,7 @@ def _parse_block(lines: List[String], start: Int, end: Int) raises -> List[Cue]:
     while seg_start < end:
         var t = -1
         for j in range(seg_start, end):
-            if lines[j].find("-->") != -1:
+            if _is_timing_line(lines[j]):
                 t = j
                 break
         if t == -1:
@@ -299,14 +317,24 @@ def _parse_block(lines: List[String], start: Int, end: Int) raises -> List[Cue]:
                 index = index * 10 + Int(b) - ord("0")
         var start_ms = 0
         var end_ms = 0
-        _parse_timing(lines[t], start_ms, end_ms)
+        # Isolate each segment's parse: a malformed timing line skips just
+        # this segment, never discarding cues already gathered from the
+        # block. `_is_timing_line` already validated `lines[t]`, so this
+        # is defense-in-depth against divergence between the two.
+        try:
+            _parse_timing(lines[t], start_ms, end_ms)
+        except:
+            seg_start = t + 1
+            continue
         # A block can hold a second (or third...) cue glued on with no
         # blank-line separator. If a later "text" line is itself a
         # timing line, that's where this cue's text ends and the next
         # cue begins — including its optional index line just before it.
+        # A bare `-->` inside cue text is not a boundary; only a line that
+        # parses as `timestamp --> timestamp` is.
         var text_end = end
         for j in range(t + 1, end):
-            if lines[j].find("-->") != -1:
+            if _is_timing_line(lines[j]):
                 if j > t + 1 and _is_all_digits(lines[j - 1]):
                     text_end = j - 1
                 else:
diff --git a/test/test_captions.mojo b/test/test_captions.mojo
index 520ba8b..b3d9fe9 100644
--- a/test/test_captions.mojo
+++ b/test/test_captions.mojo
@@ -334,6 +334,51 @@ def test_srt_glued_cues_without_blank_line() raises:
     assert_equal(caps.cues[1].text, "Second cue text.")
 
 
+def test_arrow_in_cue_text_not_a_boundary() raises:
+    """A `-->` inside cue text is prose, not a timing line: the cue must
+    survive whole rather than being split (or discarded) as a boundary."""
+    var caps = parse_captions(
+        String("1\n00:00:01,000 --> 00:00:02,000\nUse map --> filter here.\n")
+    )
+    assert_equal(len(caps.cues), 1)
+    assert_equal(caps.cues[0].text, "Use map --> filter here.")
+    # And it must round-trip through both serializers unchanged.
+    var via_srt = parse_captions(to_srt(caps))
+    assert_equal(len(via_srt.cues), 1)
+    assert_equal(via_srt.cues[0].text, "Use map --> filter here.")
+    var via_vtt = parse_captions(to_vtt(caps))
+    assert_equal(len(via_vtt.cues), 1)
+    assert_equal(via_vtt.cues[0].text, "Use map --> filter here.")
+
+
+def test_arrow_in_multiline_text_preserved() raises:
+    """A `-->` on a later text line must not truncate the cue or spawn a
+    bogus second cue; the full multi-line text is kept."""
+    var caps = parse_captions(
+        String(
+            "1\n00:00:01,000 --> 00:00:02,000\n"
+            "first line\nx --> y transform\nthird line\n"
+        )
+    )
+    assert_equal(len(caps.cues), 1)
+    assert_equal(caps.cues[0].text, "first line\nx --> y transform\nthird line")
+
+
+def test_arrow_text_still_splits_glued_real_cue() raises:
+    """Even when a cue's text holds a `-->`, a genuinely glued second cue
+    (a real timing line, no blank separator) is still split out."""
+    var caps = parse_captions(
+        String(
+            "1\n00:00:01,000 --> 00:00:02,000\nUse map --> filter here.\n"
+            "2\n00:00:03,000 --> 00:00:04,000\nSecond cue text.\n"
+        )
+    )
+    assert_equal(len(caps.cues), 2)
+    assert_equal(caps.cues[0].text, "Use map --> filter here.")
+    assert_equal(caps.cues[1].index, 2)
+    assert_equal(caps.cues[1].text, "Second cue text.")
+
+
 def test_srt_explicit_zero_index_roundtrip() raises:
     """An explicit cue number of 0 is a real index, not the "absent"
     sentinel, and must survive a to_srt round trip unchanged."""