Simplified the json schema passed to the LLM, and implemented more forgiving error handling for LLM proposed corrections

tests/test_corrections.py

@@ -20,69 +20,95 @@ def test_apply_corrections_uses_threshold_and_sorts_chronologically():
     transcript = _transcript()
     corrections = [
         CorrectionCandidate(
-            segment_index=0,
-            speaker="Eric",
-            start=10.0,
-            end=11.0,
+            segment_id=0,
             original_text="I ask Chontia.",
             corrected_text="I ask Chauntea.",
             confidence=0.8,
         )
     ]
 
-    revised = apply_corrections(transcript, corrections, confidence_threshold=0.8)
+    result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
-    assert [segment.speaker for segment in revised] == ["Mike", "Eric"]
-    assert revised[1].text == "I ask Chauntea."
+    assert [segment.speaker for segment in result.transcript] == ["Mike", "Eric"]
+    assert result.transcript[1].text == "I ask Chauntea."
+    assert result.skipped == []
 
 
 def test_apply_corrections_ignores_below_threshold():
     transcript = _transcript()
     corrections = [
         CorrectionCandidate(
-            segment_index=0,
-            speaker="Eric",
-            start=10.0,
-            end=11.0,
+            segment_id=0,
             original_text="I ask Chontia.",
             corrected_text="I ask Chauntea.",
             confidence=0.79,
         )
     ]
 
-    revised = apply_corrections(transcript, corrections, confidence_threshold=0.8)
+    result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
-    assert revised[1].text == "I ask Chontia."
+    assert result.transcript[1].text == "I ask Chontia."
+    assert result.skipped == []
 
 
-def test_apply_corrections_rejects_duplicate_targets():
+def test_apply_corrections_skips_duplicate_targets():
     transcript = _transcript()
-    correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=10.0,
-        end=11.0,
+    first = CorrectionCandidate(
+        segment_id=0,
         original_text="I ask Chontia.",
         corrected_text="I ask Chauntea.",
         confidence=0.8,
     )
+    second = CorrectionCandidate(
+        segment_id=0,
+        original_text="I ask Chontia.",
+        corrected_text="I ask Something Else.",
+        confidence=0.9,
+    )
 
-    with pytest.raises(AuditaValidationError):
-        apply_corrections(transcript, [correction, correction], confidence_threshold=0.8)
+    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
 
 
-def test_apply_corrections_rejects_mismatched_original_text():
+def test_apply_corrections_skips_mismatched_original_text():
     transcript = _transcript()
     correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=10.0,
-        end=11.0,
+        segment_id=0,
         original_text="Different text.",
         corrected_text="I ask Chauntea.",
         confidence=0.8,
     )
 
-    with pytest.raises(AuditaValidationError):
-        apply_corrections(transcript, [correction], confidence_threshold=0.8)
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
 
+    assert result.transcript[1].text == "I ask Chontia."
+    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
+
+
+def test_apply_corrections_skips_missing_segment_id():
+    transcript = _transcript()
+    correction = CorrectionCandidate(
+        segment_id=99,
+        original_text="Missing.",
+        corrected_text="Still missing.",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert [segment.text for segment in result.transcript] == ["Then Lyra.", "I ask Chontia."]
+    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_rejects_invalid_threshold():
+    with pytest.raises(AuditaValidationError):
+        apply_corrections(_transcript(), [], confidence_threshold=1.1)

tests/test_pipeline.py

@@ -1,9 +1,6 @@
-from pathlib import Path
-
-import pytest
+import json
 
 from audita.config import AuditaConfig
-from audita.errors import AuditaValidationError
 from audita.pipeline import process_transcript
 from audita.schemas import CorrectionCandidate, CorrectionSet, parse_glossary_yaml, parse_transcript_json
 
@@ -51,10 +48,7 @@ def _transcript():
 
 def test_pipeline_processes_with_fake_llm_and_cleans_work_dir(tmp_path):
     correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=0.0,
-        end=1.0,
+        segment_id=0,
         original_text="I ask Chontia.",
         corrected_text="I ask Chauntea.",
         confidence=0.95,
@@ -73,27 +67,30 @@ def test_pipeline_processes_with_fake_llm_and_cleans_work_dir(tmp_path):
     assert list((tmp_path / "work").iterdir()) == []
 
 
-def test_pipeline_preserves_work_dir_on_failure(tmp_path):
+def test_pipeline_skips_bad_correction_and_preserves_diagnostics(tmp_path):
     correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=0.0,
-        end=1.0,
+        segment_id=0,
         original_text="Different text.",
         corrected_text="I ask Chauntea.",
         confidence=0.95,
     )
     fake_client = FakeLLMClient([CorrectionSet(corrections=[correction])])
+    progress = []
 
-    with pytest.raises(AuditaValidationError):
-        process_transcript(
-            _transcript(),
-            _glossary(),
-            _config(tmp_path),
-            llm_client=fake_client,
-        )
+    revised = process_transcript(
+        _transcript(),
+        _glossary(),
+        _config(tmp_path),
+        llm_client=fake_client,
+        progress=progress.append,
+    )
 
+    assert revised[0].text == "I ask Chontia."
+    assert any("Skipping correction for segment 0" in message for message in progress)
     preserved = list((tmp_path / "work").iterdir())
     assert len(preserved) == 1
-    assert (Path(preserved[0]) / "section-0000.json").exists()
-
+    skipped_path = preserved[0] / "skipped-corrections.json"
+    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"]

tests/test_prompts.py

@@ -1,3 +1,5 @@
+import json
+
 from audita.chunking import chunk_transcript
 from audita.prompts import build_glossary_correction_messages
 from audita.schemas import parse_glossary_yaml, parse_transcript_json
@@ -32,3 +34,31 @@ 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
+
+
+def test_prompt_uses_simplified_segment_payload():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "The gestures are nearby."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Jesters"
+            category: faction
+            summary: "The Jesters are a local faction."
+        """
+    )
+    section = chunk_transcript(transcript, max_section_tokens=16000)[0]
+
+    messages = build_glossary_correction_messages(section, glossary)
+    transcript_json = messages[1]["content"].split("Transcript section:\n", maxsplit=1)[1]
+    prompt_segments = json.loads(transcript_json)
+
+    assert prompt_segments == [{"segment_id": 0, "original_text": "The gestures are nearby."}]
+    assert "speaker" not in prompt_segments[0]
+    assert "start" not in prompt_segments[0]
+    assert "end" not in prompt_segments[0]