Grammar corrections now run through a deterministic guard that blocks protected glossary names/aliases from being changed

tests/test_config.py

@@ -33,7 +33,6 @@ def test_config_uses_defaults_with_api_key():
     assert config.normalize_max_segment_gap == DEFAULT_NORMALIZE_MAX_SEGMENT_GAP
     assert config.normalize_max_segment_gap == 5.0
     assert config.normalize_ellipsis_gap == DEFAULT_NORMALIZE_ELLIPSIS_GAP
-    assert config.normalize_ellipsis_gap == 2.0
     assert config.normalize_max_segment_duration == DEFAULT_NORMALIZE_MAX_SEGMENT_DURATION
     assert config.normalize_max_segment_duration == 60.0
     assert config.normalize_max_segment_tokens == DEFAULT_NORMALIZE_MAX_SEGMENT_TOKENS

tests/test_corrections.py

@@ -171,6 +171,56 @@ def test_apply_corrections_requires_unique_match_when_configured():
     assert "more than once" in result.skipped[0].reason
 
 
+def test_apply_corrections_skips_when_guard_rejects_replacement():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Hrank moves."}
+        ]
+        """
+    )
+    correction = CorrectionCandidate(
+        id=1,
+        original_text="Hrank",
+        corrected_text="Frank",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(
+        transcript,
+        [correction],
+        confidence_threshold=0.8,
+        replacement_mode="require_unique",
+        correction_guard=lambda before, after: "protected term changed" if before != after else None,
+    )
+
+    assert result.transcript[0].text == "Hrank moves."
+    assert len(result.skipped) == 1
+    assert result.skipped[0].reason == "protected term changed"
+    assert result.skipped[0].actual_text == "Hrank moves."
+
+
+def test_apply_corrections_without_guard_remains_permissive():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Frank moves."}
+        ]
+        """
+    )
+    correction = CorrectionCandidate(
+        id=1,
+        original_text="Frank",
+        corrected_text="Hrank",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert result.transcript[0].text == "Hrank moves."
+    assert result.skipped == []
+
+
 def test_apply_corrections_skips_empty_original_text():
     transcript = _transcript()
     correction = CorrectionCandidate(

tests/test_pipeline.py

@@ -305,6 +305,59 @@ def test_grammar_stage_runs_after_glossary_and_sees_corrected_text(tmp_path):
     assert revised[0].text == "I ask Chauntea."
 
 
+def test_grammar_stage_cannot_reverse_glossary_protected_term(tmp_path):
+    transcript = parse_source_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Frank moves."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Hrank"
+            category: pc
+            summary: "Hrank is a player character."
+        """
+    )
+    glossary_correction = CorrectionCandidate(
+        id=1,
+        original_text="Frank",
+        corrected_text="Hrank",
+        confidence=0.95,
+    )
+    grammar_reversal = CorrectionCandidate(
+        id=1,
+        original_text="Hrank",
+        corrected_text="Frank",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[glossary_correction]),
+            CorrectionSet(corrections=[grammar_reversal]),
+        ]
+    )
+    progress = []
+
+    revised = process_transcript(
+        transcript,
+        glossary,
+        _config(tmp_path, grammar_max_llm_passes=1),
+        llm_client=fake_client,
+        progress=progress.append,
+    )
+
+    assert revised[0].text == "Hrank moves."
+    assert any("Skipping grammar correction for id 1" in message for message in progress)
+    run_dirs = list((tmp_path / "work").iterdir())
+    assert len(run_dirs) == 1
+    diagnostics = json.loads((run_dirs[0] / "skipped-corrections.json").read_text(encoding="utf-8"))
+    assert diagnostics["skipped_corrections"][0]["stage"] == "grammar"
+    assert "protected glossary term" in diagnostics["skipped_corrections"][0]["reason"]
+
+
 def test_grammar_stage_retries_repeated_span_and_applies_unique_retry(tmp_path):
     transcript = parse_source_transcript_json(
         """

tests/test_prompts.py

@@ -93,7 +93,9 @@ def test_grammar_prompt_limits_readability_corrections_and_protects_glossary():
     assert "homophone fixes" in prompt_text
     assert "spelling fixes" in prompt_text
     assert "Do not paraphrase" in prompt_text
-    assert "protected vocabulary" in prompt_text
+    assert "glossary names and aliases as protected spellings" in prompt_text
+    assert "Do not autocorrect, Anglicize, replace, normalize, lowercase" in prompt_text
+    assert "Preserve canonical glossary capitalization" in prompt_text
     assert "appears exactly once" in prompt_text
     assert "Do not return speaker, start, or end fields" in prompt_text
 

tests/test_protection.py

@@ -0,0 +1,52 @@
+from audita.protection import ProtectedVocabulary
+from audita.schemas import parse_glossary_yaml
+
+
+def _vocabulary():
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Hrank"
+            aliases:
+              - "Greenfield"
+            category: pc
+            summary: "Hrank Greenfield is a player character."
+        """
+    )
+    return ProtectedVocabulary.from_glossary(glossary)
+
+
+def test_protection_blocks_replacing_protected_term():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("Hrank moves.", "Frank moves.") is not None
+
+
+def test_protection_blocks_lowercasing_protected_term():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("Hrank moves.", "hrank moves.") is not None
+
+
+def test_protection_allows_canonical_capitalization():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("hrank moves.", "Hrank moves.") is None
+
+
+def test_protection_allows_punctuation_around_protected_term():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("Hrank, moves.", "Hrank. Moves.") is None
+
+
+def test_protection_does_not_match_terms_inside_larger_words():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("The shrank spell worked.", "The shrank spell works.") is None
+
+
+def test_protection_applies_to_aliases():
+    vocabulary = _vocabulary()
+
+    assert vocabulary.violation_reason("Greenfield waits.", "greenfield waits.") is not None