diff --git a/CLAUDE.md b/CLAUDE.md
index 477cfc6..e09ec7f 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -73,6 +73,7 @@ segments[]
     text: string
     cut: boolean
   cuts: TimeCut[]         [{from, to}] intra-segment ranges to skip
+  synthetic?: boolean     true when created by a > SPEAKER split in the doc; absent on raw segments
   hook?: boolean          when true, prepended as hook/teaser before main
   hookFrom?, hookTo?      clip bounds within the segment (seconds)
   cameraCues[]            explicit camera shot overrides (> CAM directives in doc)
diff --git a/remotion/types/transcript.ts b/remotion/types/transcript.ts
index 5d89568..0bc5fca 100644
--- a/remotion/types/transcript.ts
+++ b/remotion/types/transcript.ts
@@ -66,6 +66,8 @@ export type Segment = {
   graphics: GraphicsCue[];
   /** Explicit camera cut overrides — take priority over the pacing algorithm */
   cameraCues?: CameraCue[];
+  /** True when this segment was created by a > SPEAKER split in the doc; absent on raw segments */
+  synthetic?: boolean;
   /** When true, this segment is prepended to the video as a hook/teaser */
   hook?: boolean;
   /** The specific phrase within the segment used as the hook clip */
diff --git a/scripts/edit-transcript.js b/scripts/edit-transcript.js
index cdb399e..6c3a88d 100644
--- a/scripts/edit-transcript.js
+++ b/scripts/edit-transcript.js
@@ -1222,6 +1222,7 @@ function mergeDocIntoTranscript(transcript, docContent) {
       const synthSeg = {
         ...parentSeg,
         id: synthId,
+        synthetic: true,
         speaker: newSpeaker,
         tokens: tokensB,
         start: splitTime,
@@ -1823,6 +1824,22 @@ function buildPrevTokensByTdtw(tokens) {
   return map;
 }
 
+// ─── Re-injection helper ──────────────────────────────────────────────────────
+
+/**
+ * Returns segments from a prior transcript run that should be re-injected into
+ * the new transcript. Only synthetic segments (created by > SPEAKER splits) are
+ * eligible — real segments whose timestamps shifted after retranscription must
+ * not be duplicated.
+ *
+ * @param {import('./types/transcript').Segment[]} existingSegments
+ * @param {Set<number>} matchedExistingIds  IDs already matched to new sentences
+ * @returns {import('./types/transcript').Segment[]}
+ */
+export function reInjectSyntheticSegments(existingSegments, matchedExistingIds) {
+  return existingSegments.filter(s => s.synthetic && !matchedExistingIds.has(s.id));
+}
+
 // ─── Main ─────────────────────────────────────────────────────────────────────
 
 async function main() {
@@ -1927,7 +1944,10 @@ async function main() {
 
     // Re-inject synthetic segments (created by > SPEAKER splits) that were never
     // matched by findPrev — these have no corresponding raw sentence.
-    const syntheticSegs = existing.segments.filter(s => !matchedExistingIds.has(s.id));
+    // Only segments explicitly marked synthetic are re-injected; unmarked unmatched
+    // segments are real sentences from a previous run whose timestamps no longer
+    // align after a fresh transcription/alignment pass.
+    const syntheticSegs = reInjectSyntheticSegments(existing.segments, matchedExistingIds);
     if (syntheticSegs.length > 0) {
       const syntheticStarts = new Set(syntheticSegs.map(s => s.start));
       const merged = [...transcript.segments, ...syntheticSegs].sort((a, b) => a.start - b.start);
diff --git a/scripts/edit-transcript.test.js b/scripts/edit-transcript.test.js
index 945ed1f..f3c4574 100644
--- a/scripts/edit-transcript.test.js
+++ b/scripts/edit-transcript.test.js
@@ -6,19 +6,15 @@ import {
   deriveCuts,
   cleanCaptionText,
   buildSentencesVtt,
-  buildSentencesSrt,
   buildYouTubeSubtitles,
-  getSubClips,
-  getHookClips,
   resolvePhraseToTimeRange,
-  resolvePhraseToFirstTokenIndex,
   autoCutPauses,
   autoCutDisfluencies,
   rebalanceBoundaryTokens,
   buildPrevTokensByTdtw,
+  reInjectSyntheticSegments,
   WORD_DURATION_ESTIMATE,
   CUT_START_BIAS,
-  CUT_END_BIAS,
   isSpecialToken,
   isDisfluencyToken,
 } from './edit-transcript.js';
@@ -384,7 +380,6 @@ describe('applyTextPartsToTokens', () => {
   test('ignores colon-reason syntax from old docs (backwards compat)', () => {
     // Old format {um:filler} — reason is stripped, word still gets cut
     const tokens = [tok(' um', 0.1), tok(' Hello', 0.2)];
-    const result = applyTextPartsToTokens('{um:filler} Hello', tokens);
     // "um:filler" is treated as the full span — won't match token "um"
     // This is acceptable: old docs with reasons just won't cut those tokens
     // (the important thing is it doesn't crash)
@@ -1298,6 +1293,18 @@ describe('> SPEAKER split', () => {
     expect(synth.text).toMatch(/^Right/i);
   });
 
+  test('> SPEAKER split marks the created segment with synthetic: true', () => {
+    const result = mergeDocIntoTranscript(makeSplitTranscript(), doc);
+    const synth = result.segments.find(s => s.id !== 1 && s.id !== 2 && s.speaker === 'Natasha' && s.start < 20);
+    expect(synth.synthetic).toBe(true);
+  });
+
+  test('real segments do not carry synthetic: true', () => {
+    const result = mergeDocIntoTranscript(makeSplitTranscript(), doc);
+    const real = result.segments.filter(s => s.id === 1 || s.id === 2);
+    for (const s of real) expect(s.synthetic).toBeFalsy();
+  });
+
   test('no segment overlaps in output', () => {
     const result = mergeDocIntoTranscript(makeSplitTranscript(), doc);
     const active = result.segments.filter(s => !s.cut).sort((a, b) => a.start - b.start);
@@ -1323,3 +1330,32 @@ describe('> SPEAKER split', () => {
     expect(synth).toBeDefined();
   });
 });
+
+// ── reInjectSyntheticSegments ──────────────────────────────────────────────────
+
+describe('reInjectSyntheticSegments', () => {
+  const makeSeg = (id, synthetic) => ({
+    id, start: id * 10, end: id * 10 + 5, speaker: 'X', text: 'hi',
+    cut: false, tokens: [], cuts: [], graphics: [], synthetic,
+  });
+
+  test('re-injects synthetic segments whose id is not in matchedIds', () => {
+    const segs = [makeSeg(1, true), makeSeg(2, true)];
+    const result = reInjectSyntheticSegments(segs, new Set());
+    expect(result).toHaveLength(2);
+    expect(result.map(s => s.id)).toEqual([1, 2]);
+  });
+
+  test('does not re-inject real segments even when id is not in matchedIds', () => {
+    const segs = [makeSeg(1, false), makeSeg(2, undefined)];
+    const result = reInjectSyntheticSegments(segs, new Set());
+    expect(result).toHaveLength(0);
+  });
+
+  test('does not re-inject synthetic segments whose id was already matched', () => {
+    const segs = [makeSeg(1, true), makeSeg(2, true)];
+    const result = reInjectSyntheticSegments(segs, new Set([1]));
+    expect(result).toHaveLength(1);
+    expect(result[0].id).toBe(2);
+  });
+});