diff --git a/__TEST__/caption.test.js b/__TEST__/caption.test.js
index a691532..fa5c2c2 100644
--- a/__TEST__/caption.test.js
+++ b/__TEST__/caption.test.js
@@ -209,3 +209,38 @@ test("cue timing data round-trips through the data property", () => {
   expect(result.data[0].start).toBe("00:00:01.000");
   expect(flat(result.data[0].text)).toBe("Offset start.");
 });
+
+test("a missing data-d next to a non-increasing data-m never goes negative (hyperaudio-lite-editor#411)", () => {
+  // "Alone." has no data-d and the NEXT word's data-m sits before its own
+  // (out-of-order times after edits; duplicate stamps on split tokens are the
+  // milder case). The derived duration went negative, producing a cue with
+  // stop < start — dropped by browsers, and formatSeconds wrapped the negative
+  // to a ~24h timestamp.
+  buildTranscript([
+    [2000, null, "Alone."],
+    [1500, 300, "Next."],
+  ]);
+  const cues = parseVtt(caption().init("transcript", null).vtt);
+
+  expect(cues.length).toBeGreaterThanOrEqual(1);
+  for (const cue of cues) {
+    expect(cue.stop > cue.start).toBe(true);   // lexicographic works for HH:MM:SS.mmm
+    expect(cue.stop.startsWith("23:59")).toBe(false);
+  }
+});
+
+test("an inverted or zero-length cue is repaired to a readable length, not skipped (hyperaudio-lite-editor#411)", () => {
+  // A single-word cue whose word has an explicit zero duration serialized as
+  // stop == start; the timing safeguard stepped over exactly these, so the
+  // zero-length cue shipped (and browsers dropped it from the track).
+  buildTranscript([
+    [0, 500, "One."],
+    [5000, 0, "Beep."],
+  ]);
+  const cues = parseVtt(caption().init("transcript", null).vtt);
+
+  expect(cues).toHaveLength(2);
+  expect(cues[1].start).toBe("00:00:05.000");
+  // repaired: extended to the 1s minimum on-screen time
+  expect(cues[1].stop).toBe("00:00:06.000");
+});
diff --git a/js/caption.js b/js/caption.js
index 74fef10..ce3804c 100644
--- a/js/caption.js
+++ b/js/caption.js
@@ -135,6 +135,12 @@ const caption = function () {
             if (thisDuration > maxWordDuration) {
               thisDuration = maxWordDuration;
             }
+            if (thisDuration < 0) {
+              // the next word's data-m can sit at or before this word's
+              // (duplicate stamps on split tokens, or out-of-order times after
+              // edits) — a negative duration would invert the cue downstream
+              thisDuration = 0;
+            }
           } else {
             thisDuration = 5; // sensible default for the last word
           }
@@ -376,9 +382,18 @@ const caption = function () {
     function applyTimingSafeguards(caps) {
       for (let c = 0; c < caps.length; c++) {
         const start = timecodeToSeconds(caps[c].start);
-        const stop = timecodeToSeconds(caps[c].stop);
-        if (isNaN(start) || isNaN(stop) || stop <= start) {
-          continue; // skip malformed or zero-length cues
+        let stop = timecodeToSeconds(caps[c].stop);
+        if (isNaN(start) || isNaN(stop)) {
+          continue; // skip malformed cues
+        }
+
+        // An inverted or zero-length cue (stop <= start) — zero/duplicate word
+        // durations, or out-of-order times after edits — is REPAIRED rather
+        // than skipped: browsers drop such cues from a VTT track, and some SRT
+        // consumers reject the whole file. Treating its stop as the start lets
+        // the extension below give it a readable length.
+        if (stop <= start) {
+          stop = start;
         }
 
         // visible characters, counting a single space per line break
@@ -401,6 +416,12 @@ const caption = function () {
           }
         }
 
+        // never emit stop <= start, even when the next cue starts at or before
+        // this one (out-of-order cue starts) — a minimal cue beats a dropped one
+        if (desiredStop <= start) {
+          desiredStop = start + minCaptionGap;
+        }
+
         if (desiredStop > stop) {
           caps[c].stop = formatSeconds(desiredStop);
         }