Refactor to request matching/replacement substrings from LLMs, rather than requesting a complete replacement for the full original text

tests/test_corrections.py

@@ -9,7 +9,7 @@ def _transcript():
     return parse_transcript_json(
         """
         [
-          {"speaker": "Eric", "start": 10.0, "end": 11.0, "text": "I ask Chontia."},
+          {"speaker": "Eric", "start": 10.0, "end": 11.0, "text": "I ask Chontia for help."},
           {"speaker": "Mike", "start": 0.0, "end": 1.0, "text": "Then Lyra."}
         ]
         """
@@ -21,8 +21,8 @@ def test_apply_corrections_uses_threshold_and_sorts_chronologically():
     corrections = [
         CorrectionCandidate(
             segment_id=0,
-            original_text="I ask Chontia.",
-            corrected_text="I ask Chauntea.",
+            original_text="Chontia",
+            corrected_text="Chauntea",
             confidence=0.8,
         )
     ]
@@ -30,7 +30,7 @@ def test_apply_corrections_uses_threshold_and_sorts_chronologically():
     result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
     assert [segment.speaker for segment in result.transcript] == ["Mike", "Eric"]
-    assert result.transcript[1].text == "I ask Chauntea."
+    assert result.transcript[1].text == "I ask Chauntea for help."
     assert result.skipped == []
 
 
@@ -39,57 +39,55 @@ def test_apply_corrections_ignores_below_threshold():
     corrections = [
         CorrectionCandidate(
             segment_id=0,
-            original_text="I ask Chontia.",
-            corrected_text="I ask Chauntea.",
+            original_text="Chontia",
+            corrected_text="Chauntea",
             confidence=0.79,
         )
     ]
 
     result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
-    assert result.transcript[1].text == "I ask Chontia."
+    assert result.transcript[1].text == "I ask Chontia for help."
     assert result.skipped == []
 
 
-def test_apply_corrections_skips_duplicate_targets():
+def test_apply_corrections_allows_multiple_distinct_spans_in_one_segment():
     transcript = _transcript()
     first = CorrectionCandidate(
         segment_id=0,
-        original_text="I ask Chontia.",
-        corrected_text="I ask Chauntea.",
+        original_text="Chontia",
+        corrected_text="Chauntea",
         confidence=0.8,
     )
     second = CorrectionCandidate(
         segment_id=0,
-        original_text="I ask Chontia.",
-        corrected_text="I ask Something Else.",
+        original_text="help",
+        corrected_text="guidance",
         confidence=0.9,
     )
 
     result = apply_corrections(transcript, [first, second], confidence_threshold=0.8)
 
-    assert result.transcript[1].text == "I ask Chauntea."
-    assert len(result.skipped) == 1
-    assert result.skipped[0].segment_id == 0
-    assert "duplicate" in result.skipped[0].reason
+    assert result.transcript[1].text == "I ask Chauntea for guidance."
+    assert result.skipped == []
 
 
-def test_apply_corrections_skips_mismatched_original_text():
+def test_apply_corrections_skips_missing_substring():
     transcript = _transcript()
     correction = CorrectionCandidate(
         segment_id=0,
         original_text="Different text.",
-        corrected_text="I ask Chauntea.",
+        corrected_text="Chauntea",
         confidence=0.8,
     )
 
     result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
 
-    assert result.transcript[1].text == "I ask Chontia."
+    assert result.transcript[1].text == "I ask Chontia for help."
     assert len(result.skipped) == 1
     assert result.skipped[0].segment_id == 0
-    assert result.skipped[0].actual_text == "I ask Chontia."
-    assert "original_text" in result.skipped[0].reason
+    assert result.skipped[0].actual_text == "I ask Chontia for help."
+    assert "does not match any substring" in result.skipped[0].reason
 
 
 def test_apply_corrections_skips_missing_segment_id():
@@ -103,12 +101,66 @@ def test_apply_corrections_skips_missing_segment_id():
 
     result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
 
-    assert [segment.text for segment in result.transcript] == ["Then Lyra.", "I ask Chontia."]
+    assert [segment.text for segment in result.transcript] == ["Then Lyra.", "I ask Chontia for help."]
     assert len(result.skipped) == 1
     assert result.skipped[0].segment_id == 99
     assert "does not exist" in result.skipped[0].reason
 
 
+def test_apply_corrections_skips_no_op():
+    transcript = _transcript()
+    correction = CorrectionCandidate(
+        segment_id=0,
+        original_text="Chontia",
+        corrected_text="Chontia",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert result.transcript[1].text == "I ask Chontia for help."
+    assert len(result.skipped) == 1
+    assert "identical" in result.skipped[0].reason
+
+
+def test_apply_corrections_skips_ambiguous_repeated_substring():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Bane met Bane."}
+        ]
+        """
+    )
+    correction = CorrectionCandidate(
+        segment_id=0,
+        original_text="Bane",
+        corrected_text="Bain",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert result.transcript[0].text == "Bane met Bane."
+    assert len(result.skipped) == 1
+    assert "multiple times" in result.skipped[0].reason
+
+
+def test_apply_corrections_skips_empty_original_text():
+    transcript = _transcript()
+    correction = CorrectionCandidate(
+        segment_id=0,
+        original_text="",
+        corrected_text="Chauntea",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert result.transcript[1].text == "I ask Chontia for help."
+    assert len(result.skipped) == 1
+    assert "empty" in result.skipped[0].reason
+
+
 def test_apply_corrections_rejects_invalid_threshold():
     with pytest.raises(AuditaValidationError):
         apply_corrections(_transcript(), [], confidence_threshold=1.1)

tests/test_pipeline.py

@@ -49,8 +49,8 @@ def _transcript():
 def test_pipeline_processes_with_fake_llm_and_cleans_work_dir(tmp_path):
     correction = CorrectionCandidate(
         segment_id=0,
-        original_text="I ask Chontia.",
-        corrected_text="I ask Chauntea.",
+        original_text="Chontia",
+        corrected_text="Chauntea",
         confidence=0.95,
     )
     fake_client = FakeLLMClient([CorrectionSet(corrections=[correction])])
@@ -71,7 +71,7 @@ def test_pipeline_skips_bad_correction_and_preserves_diagnostics(tmp_path):
     correction = CorrectionCandidate(
         segment_id=0,
         original_text="Different text.",
-        corrected_text="I ask Chauntea.",
+        corrected_text="Chauntea",
         confidence=0.95,
     )
     fake_client = FakeLLMClient([CorrectionSet(corrections=[correction])])
@@ -93,4 +93,4 @@ def test_pipeline_skips_bad_correction_and_preserves_diagnostics(tmp_path):
     assert skipped_path.exists()
     diagnostics = json.loads(skipped_path.read_text(encoding="utf-8"))
     assert diagnostics["skipped_corrections"][0]["segment_id"] == 0
-    assert "original_text" in diagnostics["skipped_corrections"][0]["reason"]
+    assert "does not match any substring" in diagnostics["skipped_corrections"][0]["reason"]

tests/test_prompts.py

@@ -34,6 +34,9 @@ def test_prompt_requires_acoustically_plausible_transcription_errors():
     assert '"gestures" to "Jesters"' in prompt_text
     assert '"Lyra" to "Jesters"' in prompt_text
     assert "should be omitted" in prompt_text
+    assert "exact text span that needs replacement" in prompt_text
+    assert "replacement text for that span" in prompt_text
+    assert "Do not return corrections where original_text and corrected_text are identical" in prompt_text
 
 
 def test_prompt_uses_simplified_segment_payload():